Sales forecasting becomes difficult to trust when the number changes but nobody can recover what it said before. A dashboard may show an encouraging accuracy score while leaving you unsure whether the estimate was available early enough to support a real decision.
That gap matters when you are commissioning a new forecasting process. You need to distinguish a tidy reconstruction from a genuine earlier prediction, and a successful software setup from evidence that a method is useful. Without those distinctions, a supplier demonstration can answer a more convenient question than the one your business asked.
This guide explains how to specify the evidence, compare methods fairly and inspect the result. It focuses on buyer-owned acceptance tests, with clearly labelled hypothetical examples, rather than another list of forecasting techniques or a repeat of pipeline configuration.
Define what sales forecasting must help you decide
Static graphic. Enable JavaScript to explore the stages.
A forecast can look impressively precise and still answer the wrong question. Your sales director may want expected bookings this quarter, while the delivery team needs an early view of work likely to start next quarter. Both requests concern future sales. They need different dates, definitions and tests, so a single accuracy score cannot settle both.
Start with the decision that will use the estimate. Write down who needs it, when they need it and what they can still change at that point. A forecast delivered after the relevant choice has been made may describe the outcome well, but it cannot demonstrate that it helped with that earlier choice.
Give the number a clear meaning
Choose the commercial event you intend to predict before discussing software. A signed order, a deal marked closed won, an issued invoice and a received payment can occur on different dates. Your forecast brief should name one of those events, or define separate forecasts for them. Do not let a dashboard label conceal the difference.
For this guide, a booking means an accepted sales commitment counted under an agreed internal rule. That is an operational definition for the examples, not accounting advice. Your organisation must decide which evidence qualifies, which amounts belong in the measure and how finance will confirm the resulting actuals.
Imagine a hypothetical service business that expects a £60,000 contract to be agreed in March, with work starting in April. A March booking forecast could correctly include that commitment under its stated rules. It would not, by itself, establish March invoiced sales or cash receipts. Testing it against either would judge a different claim.
Put the amount basis beside the number. Decide whether it includes tax, discounts, optional work, recurring charges or the full contracted term. Where a deal has several possible values, identify which field is used and why. A supplier should be able to explain the selection without relying on the person who originally built the report.
Currency needs the same treatment. State the reporting currency, the conversion rule and when any exchange rate is fixed. Otherwise a change in the reported total might reflect a rate update rather than a different commercial expectation. Keep the original currency and amount available so the conversion can be checked.
Set the decision date and the target period
Two dates belong in every forecast brief: the last information cutoff and the period being predicted. For example, you might need an estimate prepared at the end of January for bookings during April. The test must preserve that January view. An updated estimate produced during April belongs to a different forecast horizon.
Use calendar boundaries that match the business question. If your reporting periods do not align with calendar months, make that explicit. Write down the relevant time zone as well as the date. A record arriving after the cutoff should not be admitted because another system displays the timestamp in a different local time.
Longer and shorter horizons should have separate evaluation rows. It may be reasonable for an estimate to change as new evidence arrives. That does not make the earlier number disappear. Retaining both allows you to see whether the system is useful early enough, rather than only accurate shortly before the result becomes obvious.
The horizon also affects the information you can expect to exist. A forecast for imminent bookings might use named opportunities. A more distant view may need an explicit allowance for opportunities not yet created. Ask how that allowance is estimated and how it will be distinguished from the current pipeline.
Agree the population before seeing the result
Specify which business units, products, territories and transaction types are in scope. Renewals, expansions and new customers may have different commercial patterns. You can still combine them for a board view, but the underlying definition must prevent the same commitment appearing in more than one component.
Choose an owner for those boundaries. A sales manager may know whether an opportunity belongs in the current forecast, while finance knows how the eventual outcome will be reported. Resolve disagreements in the brief, not after the supplier presents a favourable score. A forecast cannot be evaluated consistently while its target keeps changing.
A useful commissioning brief contains the following items:
- The decision the forecast is intended to support, including its owner and decision date.
- The precise event and amount being predicted, with the actuals source named.
- The target periods, forecast horizons, time zone and reporting currency.
- The included population, exclusions and treatment of cancellations or corrections.
- The simple comparison method and evidence needed to judge whether a new approach helps.
Keep this short enough to use in a meeting. Its purpose is to expose disagreements before they become formulas. If two readers interpret “April sales” differently, more detail in the model specification will not repair that initial ambiguity. Ask each person to explain the same example transaction using the brief.
Turn the brief into a practical acceptance question
A weak requirement says that the forecast must be accurate. A stronger one says that a named user can retrieve the estimate available at a specified cutoff, identify its population and reproduce its comparison with the agreed outcome. That creates something you can test even before enough future periods exist to judge predictive performance.
Separate the cost of a miss from the size of a miss. Overestimating a particular service line may affect a different decision from underestimating another. Record those consequences in plain language. You can then choose evaluation measures that reveal the mistakes you care about, rather than accepting whichever score the software makes easiest to display.
GRO’s sales pipeline and forecasting service includes work on forecast views and comparison frameworks. For a buying brief, the useful starting point is this definition of the decision and its evidence. Stage configuration is a separate task, covered in the HubSpot pipeline setup guide, rather than something to rebuild inside every forecasting exercise.
Before moving on, try a simple rejection test. If the proposed output cannot tell you its cutoff, target period and amount basis without another spreadsheet or an oral explanation, it is not yet a well-defined forecast. Resolve those missing pieces before discussing whether a more complicated method is worth commissioning.
Freeze the evidence before the result is known
Static graphic. Enable JavaScript to explore the stages.
Current CRM data is useful for running today’s business. It is not automatically a record of what you knew last month. Deal values change, expected dates move and missing details get filled in. If an evaluation uses those revised records as though they existed earlier, it can credit the forecast with information it never actually had.
A frozen snapshot is a dated record of the forecast and the information admitted to it. “Frozen” means the original evidence remains retrievable when later updates arrive. It does not mean the operational CRM must stop changing. Your team should continue correcting records while the evaluation preserves a clearly labelled historical view.
Keep three different times separate
For a material input, distinguish when the business event happened, when the system recorded it and when the forecasting process could use it. Those times may coincide, but the process should not assume that they do. An integration delay or a backdated entry can put several days between the first and the last.
Consider a hypothetical order agreed on 29 March but entered into the CRM on 3 April. A forecast frozen on 31 March could only use that order if its information was genuinely available through an admitted source by then. The later entry does not prove that the forecasting process had access to it at the cutoff.
The actuals record can still count the order in March under the agreed outcome rule. That is a separate decision. Keep a note showing that the outcome belongs to March while the CRM record arrived in April. This lets the review distinguish a late information problem from a calculation or commercial judgement problem.
Ask the supplier to demonstrate this with a controlled example. Provide an input with an earlier event date and a later recorded date, then reconstruct the earlier forecast. The late input should follow the documented availability rule. If the system silently uses its final state, the demonstration has failed an evidence test, whatever number it produces.
Save enough to reconstruct the estimate
A screenshot preserves what a user saw, but often leaves out the information needed to reproduce it. A useful snapshot should identify its input extract, the included records, the calculation version and any adjustments. The appropriate format depends on the system, but the evidence should not depend on a single employee remembering how it worked.
Record a unique snapshot identifier and keep it with every exported result. Include the cutoff, creation timestamp, target period, horizon, currency and population definition. Where records are excluded, preserve the reason or a reproducible exclusion rule. Otherwise a missing opportunity can become impossible to explain once the current CRM view has changed.
Retain the input values used by the method alongside links to live records. A link is helpful for investigation, yet it may now show a revised amount or stage. The saved evidence must make clear which values were used then and which values are being displayed now.
For a calculated method, save its configuration and effective version. For a manager submission, save the submitted figure, author, submission time and stated assumptions. If a forecast combines both, preserve the components. A single final total cannot show whether a later difference arose from the model, a manual adjustment or an omitted input.
Do not include unnecessary personal information merely because it is present in the source export. Identify the fields required for the forecast and its audit trail, then apply the organisation’s existing access and retention arrangements. This guide is about commercial estimates, not creating another unrestricted store of customer or employee information.
Treat a revised forecast as another observation
Suppose a manager submits an estimate on Monday and revises it on Thursday after a material order change. Both observations can be useful. Keep the Monday snapshot, add the Thursday version and label the reason. Replacing Monday’s value with Thursday’s makes it impossible to judge the original estimate on its own terms.
A revision log should say whether the change reflects new evidence, a corrected input or a different judgement. These categories help the review ask the right question. New evidence arriving on time may be a normal part of the process. A correction to a calculation might indicate that previously released forecasts need a separate explanation.
Use clear names in the interface and exports: original submission, latest submission and final actual are different objects. Avoid a column called “forecast” whose meaning changes depending on when the report is opened. If the same report can show several versions, display the selection prominently rather than hiding it in a filter panel.
An export test can reveal the problem quickly. Retrieve one historical period twice, once before and once after a controlled record update. The original snapshot should remain identical. A revised view may change, but its identifier and status must distinguish it from the preserved original. Record both results in the acceptance evidence.
Admit incomplete history honestly
Sometimes the organisation has retained only current values and final outcomes. In that case, an apparently detailed historical forecast may be a reconstruction based on assumptions. It can help explore a method, but it should not be presented as proof of what would have been predicted at the time.
Ask which historical fields are genuinely recoverable and which are inferred. Field history may cover some values but not related records, imported data or manual assumptions. The answer needs to be specific to your account and retention history. A general statement that the platform has an audit trail is not enough.
Where the historical record is incomplete, begin collecting prospective snapshots. Run the proposed method alongside the existing process without allowing it to replace the operational forecast immediately. This creates a clean evidence set for later comparison. The length of that observation period should reflect the decisions and outcome cycle you need to assess.
Useful structure matters here. The CRM data model service is relevant when fields and relationships cannot preserve the distinctions your forecast requires. Agree the evidence model first. You can then ask a supplier to show that a result is reconstructable, rather than accepting a polished chart that cannot explain its own history.
Choose a baseline your data can support
Static graphic. Enable JavaScript to explore the stages.
A proposed forecasting method needs a credible comparison. Without one, a supplier can show that its estimates sometimes resemble actuals without demonstrating that the additional work is useful. A baseline is a clearly specified reference method that you can calculate consistently. It provides a starting point for a fair test, not a claim about the best method for every business.
Keep the baseline simple enough to explain to someone who did not build it. You might test a previous comparable period, an agreed historical average or the organisation’s existing submitted forecast. The choice depends on the decision and the records available. Write down the rule before using the results to choose a preferred answer.
Start with the information you really have
If you have reliable monthly outcomes but no trustworthy historical opportunity snapshots, a method based on those outcomes may be testable retrospectively. A historical pipeline model may not be. This is a reason to distinguish the evidence available for each candidate, rather than giving every method the same appearance of certainty.
With reliable snapshots, you can test an opportunity-based candidate against the same target periods. Record how it converts the admitted opportunities into an estimate. A stage-weighted calculation, for example, requires an explicit probability rule and amount basis. Do not assume that default percentages represent observed probabilities for your population or the horizon being evaluated.
Manager submissions are another candidate, not an untouchable truth. They may contain information that is absent from structured fields, such as a documented customer dependency. Preserve them as submitted so they can be compared with calculated alternatives. Keeping the methods separate first makes it easier to understand what a later combination contributes.
Where the records cannot support a proposed method, ask for a narrower initial test. A supplier should be able to describe the missing information and how it would be collected. Requiring a complicated model before that evidence exists can leave you paying for a process whose historical claims cannot be checked.
| Available evidence | Candidate worth testing | Evidence still needed |
|---|---|---|
| Consistent period outcomes | A simple outcome-based reference | Comparable scope and a fixed calculation rule |
| Historical as-of opportunity snapshots | A defined pipeline estimate | Input availability, amount basis and probability rule |
| Dated manager submissions | The submitted operational view | Cutoff, scope and recorded adjustments |
| Current records only | A prospective shadow comparison | New frozen snapshots and later agreed actuals |
Make simple rules genuinely reproducible
A rule such as “use the recent average” still leaves several choices open. Which periods are recent? Are they complete? What happens if a period is missing? Does the calculation include an exceptional transaction? Define these details so another person can reproduce the same result without choosing their own interpretation.
For an original hypothetical example, suppose the last three complete comparable months recorded bookings of £84,000, £96,000 and £105,000. Their arithmetic mean is £95,000. If your agreed baseline uses that three-month mean, £95,000 is its next estimate before any other rule is applied. It is an illustration of a reproducible rule, not a forecast for a real business.
Now suppose one month is unavailable. Do not silently divide the remaining sum by three or substitute a zero. Either apply an agreed missing-data rule or mark the baseline unavailable for that origin. The decision changes the comparison population, so it belongs in the test record rather than in an undocumented spreadsheet repair.
A same-period-last-year reference can be considered where its comparison makes business sense. It should not be treated as evidence of seasonality just because twelve months have passed. Changes in the product mix, territory, reporting definition or operating model may make the earlier period a poor reference. Keep those differences visible.
Compare like with like before adding complexity
Every candidate must forecast the same outcome at the same cutoff. Giving one method information from the following week makes the result harder to interpret, even if both estimates carry the same target-month label. Check the admitted inputs as carefully as the formula.
Apply the same population and currency rules to each method. If a proposed system excludes difficult opportunities while the baseline includes them, a smaller error may describe a different task. Require a list of eligible forecast cases and a record of any method-specific failures. An unavailable forecast is a result to investigate, not a row to discard unnoticed.
Where a new business line has little history, label the limitation rather than blending it invisibly into an established line. You may decide to use a different reference for that line or keep its estimate judgement-led for now. The chosen fallback should remain explicit in the aggregate view so buyers know where evidence is strongest and weakest.
The CRM data cleansing guide covers the separate job of making source records dependable. In this evaluation, the practical question is narrower: can you identify which records were usable at each cutoff and apply the same eligibility rules to every candidate? Cleaning the current database does not recreate missing historical information.
Set a reason to pay for the extra method
A more involved approach may require additional fields, maintenance, training or specialist support. Ask what improvement would justify that work for your use case. The answer might involve a particular horizon, a costly type of miss or a more usable explanation. It should not rest only on the appearance of sophistication.
Define the comparison criteria before the final evaluation. You can require both a useful error result and an operational capability, such as reconstructing a forecast after an administrator leaves. These are different tests. Passing one should not quietly compensate for failing the other unless that tradeoff was deliberately agreed.
Keep the baseline after a new method is introduced. It remains a reference when data changes, a supplier updates its model or an input becomes unavailable. A fallback is easier to trust when people already understand its rule and limitations. The goal is a forecast you can evaluate and operate, rather than a system whose performance can only be described by its seller.
Test methods in the order time actually passes
Static graphic. Enable JavaScript to explore the stages.
A method can describe completed history well because it has already seen the answers. Your buying decision concerns what it can estimate before those answers are known. The evaluation therefore needs to recreate the order in which information becomes available, rather than simply dividing a finished spreadsheet into convenient groups of rows.
The forecast evaluation discussion in Forecasting: Principles and Practice distinguishes genuine forecasts of held-out observations from fitted residuals. It also describes rolling-origin evaluation using earlier information to predict later observations. For a buyer, the practical requirement is to identify each forecast origin and the future horizon being tested.
Build a calendar of forecast origins
Choose a sequence of cutoffs that represents how the business will use the forecast. At each cutoff, record the eligible information, produce the estimate and reserve the later outcome for scoring. Move to the next cutoff only under the agreed updating rule. This creates an evaluation that can be explained as a series of dated decisions.
For example, a hypothetical monthly process might freeze a forecast at the end of each month for the month two calendar months ahead. A January cutoff predicts March, a February cutoff predicts April and a March cutoff predicts May. These are matched lead times. A last-day-of-March estimate of March is not another observation of the same task.
Document what happens between those cutoffs. If the method is retrained, say when and with which newly available outcomes. If it remains fixed, preserve that version. A test that changes the updating rule midway can be difficult to interpret, even when every individual calculation is technically reproducible.
The evaluation calendar should also allow for the delay in agreeing actuals. If a completed month is not reconciled until a later date, a method retrained before that date cannot use the final reconciled values as if they were already available. Record the version of the outcome data admitted to each update.
Separate development from the final comparison
Use one part of the available history to develop candidate rules and another to compare choices during development. Keep a later evaluation period untouched until the choice and its settings are fixed. Once a supplier has repeatedly adjusted a method in response to those final results, the period no longer serves the same purpose.
That boundary is an agreement about behaviour as much as a division of data. Ask who can inspect the reserved outcomes, when they can inspect them and how any change will be recorded. If a defect requires reopening the method, describe the new evaluation plan rather than continuing to call the old result untouched.
You may discover that there is too little usable history for every proposed partition. Say so. A shorter development exercise followed by prospective observation may be more honest than a long historical chart built from reconstructed inputs. The test design should reflect the evidence you possess, not a preferred slide layout.
Do not hide a weak final comparison behind a stronger development result. Both can be reported, with their roles labelled. The final record should show which evidence informed the method choice and which evidence was used only after that choice. This allows a buyer to understand how much of the apparent success was already known during development.
Choose the unit of evaluation carefully
A spreadsheet may contain thousands of opportunities but only a small number of completed forecast periods. Those are not interchangeable counts. If your decision uses a monthly business total, the evaluation needs to show the number of monthly forecast outcomes separately from the number of deals in the source data.
The same opportunity can appear in several successive snapshots. That may be correct for operational forecasting, but it does not make every appearance a separate independent commercial event. Keep identifiers that allow the reviewer to recognise repeated opportunities and understand how those appearances contribute to period-level scores.
Avoid moving later records into an earlier training set merely to balance its size. Calendar order should follow the available-at evidence. When a source has irregular arrivals, first define the actual business cutoffs and target periods. A software splitter cannot infer those commercial meanings from a row number.
The scikit-learn TimeSeriesSplit documentation describes ordered train and test splits and notes the equal-spacing condition for comparable fold durations. Its optional gap excludes observations between training and testing. That is a technical facility, not proof that irregular CRM events have been turned into a valid sales evaluation. Ask the implementer how its configuration matches your calendar and information delays.
Use a shadow run when history cannot answer
A prospective shadow run produces dated predictions alongside the current process, then waits for the agreed outcomes. Decide in advance which people can see the experimental forecast and whether it will influence operational decisions. If its visibility changes how the business acts, record that context rather than pretending the observation was unaffected.
Keep the existing forecast and the candidate on the same schedule. A missed submission should remain visible. So should a late export, a failed calculation or a manual rescue. These events help you assess whether the proposed service can run reliably, even before you have enough outcomes to compare its predictive usefulness.
At each origin, save the candidate version and the original baseline. Do not recreate the baseline afterwards from a cleaned database. A fair prospective comparison requires both methods to live within the same information boundary, including awkward periods when the data is incomplete.
Agree how the shadow run will end. The decision could be to continue observing, adopt a bounded use, change the method or stop the project. Tie that decision to the previously specified evidence and operational requirements. A calendar deadline alone does not make uncertain predictive evidence conclusive.
Ask for the evaluation to be replayed
Choose a completed test origin and ask someone other than the original builder to reproduce it from the saved pack. They should identify the admitted inputs, run the stated version and obtain the same estimate within the documented numerical tolerance. Any unexplained difference is a practical issue to resolve before acceptance.
Then choose an origin with a known late input or missing submission. A useful replay should reproduce the exception as well as the successful calculation. This is where an acceptance exercise becomes more informative than a demonstration: you learn whether the historical boundary survives an awkward case. A consistent presentation on a favourable date offers less evidence.
Measure forecast error without hiding the misses
Static graphic. Enable JavaScript to explore the stages.
An accuracy label can conceal several different calculations. Before accepting one, ask for the forecast value, the actual value and the exact formula behind it. The result should be reproducible outside the presentation layer. It also needs units and a stated population, otherwise two apparently similar percentages may describe different kinds of error.
In this guide, signed error means forecast minus actual. A positive result is an overprediction; a negative result is an underprediction. Other systems may use the opposite sign, so display the convention. Absolute error ignores that direction and records the size of the miss. Keeping both helps distinguish a persistent tendency from overall distance.
Work through an original example
The following figures are hypothetical and exist only to show the arithmetic. Each of five periods has actual bookings of £100,000. Their forecasts differ. No figure represents GRO performance, a client result or a benchmark that another business should expect to achieve.
| Period | Forecast | Actual | Signed error | Absolute error |
|---|---|---|---|---|
| A | £120,000 | £100,000 | +£20,000 | £20,000 |
| B | £80,000 | £100,000 | -£20,000 | £20,000 |
| C | £110,000 | £100,000 | +£10,000 | £10,000 |
| D | £90,000 | £100,000 | -£10,000 | £10,000 |
| E | £125,000 | £100,000 | +£25,000 | £25,000 |
The signed errors add to £25,000. Dividing by five gives a mean signed error of +£5,000. That shows the average direction in this small example. It does not mean the typical miss was only £5,000, because overpredictions and underpredictions offset one another in the signed sum.
The absolute errors add to £85,000, giving a mean absolute error, or MAE, of £17,000. To calculate root mean squared error, or RMSE, square each signed error, average those squares and take the square root. Here the result is about £18,028. Both measures return to the original money units; RMSE gives larger misses more influence through squaring.
These are different descriptions of the same five observations. None changes the underlying figures. A buyer can inspect the table and see that the final period had the largest miss, while the first four contained opposing errors. Keeping the observations available prevents a summary measure from becoming the only version of the evidence.
State the denominator for a percentage
A weighted absolute percentage calculation can be defined as total absolute error divided by total actual value. In this positive-actuals example, £85,000 divided by £500,000 is 17%. State that definition explicitly if you call it WAPE. The calculation weights larger actual-value periods differently from averaging their individual percentage errors. It is undefined when the chosen total denominator is zero.
Here every actual is equal, so the mean of the five absolute percentage errors is also 17%. That agreement follows from this deliberately simple example. It should not be assumed when actual values differ. A report needs to say whether it averages percentages or divides aggregated amounts, and keep the underlying denominator visible.
MAPE is undefined for an observation with zero actual value and can be unstable near zero. Do not silently replace zero with an arbitrary small number to make a score appear. Specify how zero-outcome periods are reported, and retain an absolute-money view so those periods remain visible rather than disappearing from the evaluation.
Refunds, reversals or negative net outcomes need an agreed treatment too. A signed total can cancel positive and negative amounts in a denominator, producing a percentage that is hard to interpret. Decide whether those movements belong in the forecast target, a separate measure or a documented alternative calculation. Do not improvise the rule after seeing which version looks better.
Keep direction and magnitude together
A method can have a modest average signed error while still making large opposing mistakes. A consistent small overprediction may produce a clear directional bias without the largest absolute misses. Show both views, then relate them to the decision. There is no need to compress every property into one supposedly definitive score.
The worked example also illustrates why summing a forecast across several periods can hide a poor monthly experience. A total may come closer to the accumulated actual than many individual estimates did. If your decisions occur monthly, evaluate those monthly forecasts as well as any longer-period total.
Publish the observation count beside every summary. Identify missing forecasts and missing actuals separately. If a method failed to produce a result, dropping that row can make its reported performance look cleaner than the operational experience. Agree a failure treatment and show the number of affected cases even if the statistical score uses only completed pairs.
Check rounding at the end of the calculation. Store enough precision to reproduce the score, then format a readable result. A series of rounded intermediate values can create differences between a supplier’s report and your independent calculation. The acceptance pack should explain the tolerance rather than dismissing every mismatch as rounding.
Choose a score for the decision, not the sales pitch
Ask which mistakes would make the forecast unhelpful for its intended use. You may care about a money error at a particular horizon, an ongoing tendency to overpredict or performance within one service line. A single aggregate score should not prevent you from examining those cases.
Record the preferred measures before the final comparison and show the alternatives that help interpret them. If a supplier changes its headline measure, ask why and retain the original result alongside it. A different formula may be legitimate, but it should not rewrite the agreement under which the method was evaluated.
For reporting work, the HubSpot reporting service is relevant to making definitions and calculations visible. The buying requirement is straightforward: the displayed score must trace back to named forecast and actual records. A reader should be able to understand a bad period, not merely admire a good average.
Separate timing mistakes from value mistakes
Static graphic. Enable JavaScript to explore the stages.
A forecast can miss a month even when the expected sale eventually arrives at the expected value. It can also get the month right but the amount wrong. Treating every difference as one undifferentiated accuracy problem makes it harder to decide whether the useful response concerns timing evidence, value definitions, missing opportunities or the method itself.
Create an explanation alongside the score, while keeping the original score intact. The explanation should reconcile a frozen prediction with its later outcome. It is not permission to alter the historical forecast until it matches what happened. Your aim is to understand the miss without erasing the evidence that a miss occurred.
Follow a booking across the period boundary
Take a hypothetical £60,000 booking expected in March that is instead agreed in April. Assume the amount remains £60,000 and the agreed target is bookings by calendar month. The timing moved, even though the eventual value did not. March and April therefore need different explanations from a case where the order disappeared entirely.
If that booking was included at its full £60,000 value in the March forecast, it contributes a £60,000 shortfall against that March expectation when it does not arrive. Its April arrival may then be an additional outcome relative to an April snapshot that excluded it. Keep the relevant snapshots separate; you are explaining two dated claims, not transferring one score between months.
A combined March-and-April total can conceal this timing error. That combined view may be useful for another purpose, but it cannot demonstrate that the March estimate supported a March decision. Retain the monthly evidence and explain why the longer-period view answers a different question.
The diagram uses a single full-value booking to make the timing relationship clear. A weighted forecast needs different arithmetic. If the frozen method assigned a £15,000 expected contribution to that £60,000 opportunity, its absence removes that £15,000 contribution from the forecast comparison, not automatically the full nominal deal value.
Build an explanation that adds up
Start with the frozen forecast total and the agreed actual total. The difference between them is the amount to explain. Choose a consistent direction for the bridge, such as moving from forecast to actual, and use it throughout. This avoids a report that mixes positive forecast error with positive actual movement without warning.
For an original hypothetical bridge, suppose the forecast is £180,000 and the actual is £164,000. The net movement from forecast to actual is minus £16,000. A documented explanation could include minus £24,000 from delayed expected contribution, plus £12,000 from an unanticipated booking, minus £7,000 from lower realised value and plus £3,000 from a confirmed data correction.
Those movements sum to minus £16,000: the bridge reaches £164,000. The categories are illustrative, not standard accounting classifications. Their usefulness depends on the evidence and on a rule that assigns each movement once. The arithmetic alone does not establish that the reasons are accurate.
Keep an unresolved residual when the bridge is incomplete. A visible unexplained balance is more useful than forcing every difference into an “other” category with no supporting detail. Assign someone to investigate it and retain the result of that investigation. A zero residual should mean the explanation reconciles, not that the numbers were adjusted to make it disappear.
Prevent the same movement appearing twice
A deal can change date and value at the same time. Decide how the bridge treats that combination before adding the movements. One approach is to apply an agreed sequence, with each step calculated from the state left by the previous step. Another is to use an explicit mutually exclusive classification for the purpose of that report.
Whatever approach you choose, preserve the rule with the bridge version. Two reviewers should be able to follow the same record and reach the same categories and amounts. If one assigns the full change to timing while another assigns it to value, the explanation needs a clearer convention rather than a more decorative chart.
Be cautious about assigning reasons solely from the final CRM fields. A later close date shows that the recorded date changed. It does not necessarily prove why the customer delayed, when the delay became known or whether the forecast owner could reasonably have acted on it. Distinguish the observed data movement from a supported commercial explanation.
For unexpected bookings, check whether the opportunity was genuinely absent at the cutoff or merely excluded by a filter. Those cases point to different questions. One concerns the method’s allowance for new business; the other may concern the defined population or a technical omission. Keep the evidence needed to make that distinction.
Preserve corrections without rewriting the original result
Actuals can change after the first report is released. An order may be corrected, a duplicate removed or a reporting classification revised. Keep the originally released actual and the corrected actual, with their timestamps and reason. Otherwise a reader cannot explain why last month’s published accuracy score differs from today’s version.
Choose which actuals version the formal evaluation uses and when it becomes final for that purpose. You can still publish a restated view later, provided it is labelled. The rule should apply consistently to both the new method and its baseline, rather than restating only the comparison that benefits one of them.
A movement report is also a useful conversation starter with sales and finance. Ask whether the reason categories correspond to actions someone can take: improving a recorded dependency, correcting an amount basis or repairing a missing extract. Avoid turning every unexplained miss into a judgement about an individual salesperson.
Where the underlying object relationships need work, refer back to the CRM data model rather than adding more spreadsheet exceptions. The forecast review should leave you with a clearer account of what changed and what remains uncertain. Its value lies in a traceable explanation that helps improve the process, not in making the original miss look smaller.
Check where the headline average breaks down
Static graphic. Enable JavaScript to explore the stages.
A single total is convenient for a meeting, but it can hide the exact distinction your decision depends on. Before relying on the overall result, inspect the horizons, periods and commercial groups that matter to the forecast brief. The purpose is to reveal limitations, not to search for a favourable slice after the main comparison disappoints.
Agree the important breakdowns before the final evaluation. Keep that original list in the acceptance pack and label any later exploratory analysis separately. A surprising pattern may deserve investigation, but a filter chosen after seeing the result should not quietly become the new definition of success.
Look for offsetting errors
Suppose two hypothetical service lines each record actual bookings of £100,000. The forecast predicts £140,000 for line A and £60,000 for line B. The aggregate forecast and actual are both £200,000. At that level the error is zero, even though each line is wrong by £40,000.
The decision determines how much that matters. If the two service lines need different delivery resources, their opposing misses may be important. If the only intended use is a combined total, the aggregate result still deserves to be shown. Display both views and explain which one the acceptance criteria address.
Do not average those two absolute misses into the aggregate total and call it the same metric. The £40,000 average absolute error across the two lines describes the line-level estimates. Zero aggregate error describes the combined estimate. Both calculations are correct for their stated units, but they answer different questions.
A useful report lets the reader move from the total to its components without changing definitions. If the component values do not add back to the displayed total, explain the adjustment or difference in population. Otherwise the breakdown introduces another reconciliation problem just as you are trying to understand the first one.
Keep horizons apart
A method that is useful shortly before the end of a period may be less useful at the earlier point when a decision must be made. Show results for the forecast horizons specified in the brief. Do not pool them into a single average that can be dominated by the easiest or most frequent observations.
For each horizon, show the number of completed forecast periods and their dates. If the longer-horizon test covers fewer periods, make that difference visible. A side-by-side chart should not imply equal evidence merely because the columns have the same width.
Check how repeated updates are counted. A team that submits every day can create many more rows than a team that submits weekly. If the decision requires one named cutoff per month, select that cutoff consistently. Averaging every available update may describe submission behaviour as well as predictive performance.
Late submissions deserve their own status. An accurate estimate delivered after the deadline may be unavailable for the intended decision. You can score it for analysis, but separate that score from the result of the on-time process. Otherwise a method can look operationally reliable by quietly arriving after the difficult part has passed.
Expose concentration without inventing a rule
When a few opportunities account for much of the forecast value, one outcome can move the total substantially. Show the contribution of the largest items and the effect of their timing assumptions. Avoid an arbitrary threshold presented as a universal industry standard. The relevant concentration depends on the business and the decision.
For a concentrated forecast, ask for a clearly labelled sensitivity view. What changes if the largest expected booking arrives in a different period, or at a different amount? Keep those scenarios separate from the official forecast. They help the reader understand a dependency without pretending that every alternative has a measured probability.
The same caution applies to a segment with little history. A small number of successful observations does not settle how the method will behave in every future condition. Report the counts and the observed variation. If the evidence is too thin for a strong comparison, retain that limitation rather than borrowing confidence from the larger overall population.
Avoid selecting a minimum sample size simply because it sounds reassuring. Ask the analyst to explain what can and cannot be concluded from the available periods, including their similarity and any changes in scope. The acceptance decision may need more observation or a narrower use, not a more emphatic label on the same data.
Track changes in the mix
An aggregate score can move because the work being forecast has changed. New products, different contract sizes or a shift between renewals and new business can alter the mix of cases. Keep those changes beside the comparison so a reader can distinguish a changed task from a changed forecasting method.
Suppose a difficult new service line is added midway through an evaluation. Do not remove it from the result without saying so. Show the agreed overall view and, if useful, a separate comparable-scope analysis. Explain which decision each view supports and retain the original population definition.
A comparable-scope analysis needs its own written rule. Define the included cases and apply that rule to every method and period in the comparison. If exclusions are made by inspecting individual outcomes, label the exercise exploratory. It should not replace the original acceptance score without a deliberate review of the criteria.
Record changes to sales processes that affect the inputs as well. A field becoming mandatory can improve completeness while changing what a historical comparison means. A renamed category may hide a new classification rule. The forecast record needs enough context to show when these changes occurred and which versions used them.
Make the breakdown useful to a decision-maker
Ask for a short explanation of the largest limitations, not an endless dashboard of every available filter. Name the affected decision, the evidence behind the limitation and the proposed next test. A reader should leave knowing where the forecast can be used cautiously and where the evidence remains too uncertain.
The sales pipeline and forecasting service is a relevant place to frame that comparison work. Keep the brief anchored to your chosen horizons and business groups. The strongest headline is the one that remains understandable when its components are inspected, including the cases that make the overall result less flattering.
Use ranges and judgement without false confidence
Static graphic. Enable JavaScript to explore the stages.
A point estimate is easy to circulate, but a buying decision may also need to understand what could change it. Ranges and management judgement can help explain uncertainty when their meanings are clear. They become misleading when an assumption-led scenario is presented as a measured probability, or when a manual adjustment overwrites the evidence needed to judge it.
Give every alternative view a name and a purpose. A planning scenario, a statistical prediction interval and a manager’s submitted estimate are different objects. They can appear in the same pack, but the reader should not need to infer which kind of statement a shaded band or a “likely” label represents.
Distinguish an assumption from a probability
A scenario answers a conditional question: what would the result look like under a stated set of assumptions? For example, you might move a named booking into a later period or remove a dependency that has not been confirmed. The result can be useful without claiming that the assumption has a known chance of occurring.
Write those assumptions beside the scenario. If several change at once, make that visible. An optimistic scenario that combines earlier timing, higher values and extra opportunities should not be mistaken for a small adjustment to a single input. Its meaning depends on all of those conditions together.
A prediction interval, by contrast, is associated with a specified probability under the forecasting method. Ask how that interval is produced and what assumptions it requires. A low and high number chosen in a meeting should not acquire a confidence percentage simply because the software can display one.
Avoid presenting a range as a promise that the future outcome will stay inside it. The acceptance discussion should concern how the range was constructed, how it behaved on comparable later outcomes and whether its width is useful for the decision. A range can be honest about uncertainty without providing a guarantee.
Check coverage and width together
If a supplier supplies probability-based intervals, ask it to state the intended coverage before evaluation. Then count how often later actuals fall inside the saved intervals, using a consistent boundary rule. Keep the original intervals, rather than recalculating them after the actuals are known.
For an original hypothetical example, suppose sixteen of twenty completed outcomes lie inside intervals that were labelled with an 80% target. The observed coverage in those twenty cases is 80%. That arithmetic does not establish that future coverage will be exactly 80%, or that twenty cases are enough for every conclusion you might want to draw.
Show interval widths beside coverage. A very broad range can include many outcomes while providing little help with a bounded decision. A narrow range may look decisive but exclude too many later outcomes. The buyer needs both pieces of evidence and the underlying periods, not a coverage percentage detached from its practical use.
Keep horizons and populations consistent when reviewing interval behaviour. If a report combines short-horizon and distant forecasts, its coverage figure may hide different experiences. Apply the same pre-agreed breakdown principles used for point forecasts, and state where the number of observations limits interpretation.
Preserve the forecast before an override
A manager may have information that the model does not represent. Rather than forbidding all judgement or accepting every change, record the original estimate, the adjusted estimate and the reason. Include the adjustment time and the evidence available then. This creates a testable record without claiming that every useful insight can be reduced to a field.
Ask for a concrete reason rather than a mood. “The buyer confirmed the approval meeting moved into next month” is a different kind of input from “the number feels low”. You can investigate the first through the relevant evidence. The second may still be a judgement, but it should be labelled as such rather than dressed as a confirmed event.
Prevent double counting. If an updated model input already reflects the delayed approval, a separate manual reduction for the same delay may apply the adjustment twice. The operating record should show whether an override changes an input, changes the final total or adds a separate component. Its calculation must be clear enough to inspect.
Set an expiry or review condition where an adjustment relies on a temporary circumstance. An override should not silently persist through later periods after its original reason no longer applies. Keep the history of its removal as well as its creation, so the final view can explain how the judgement evolved.
Compare adjusted and unadjusted versions fairly
Score the before-and-after estimates on the same completed cases. Include adjustments that made the error larger as well as those that reduced it. If only the successful interventions are retained, the review cannot tell you how the adjustment process behaved overall.
A favourable comparison is evidence about those recorded predictions, not automatic proof that the judgement caused a better commercial outcome. The manager may have acted on unusually difficult cases or had information unavailable elsewhere. State that context and avoid turning a simple before-and-after score into a broader causal claim.
Keep commercial targets separate from expected outcomes. A target can express ambition; a forecast describes an estimate under its stated rules. If a meeting changes the forecast solely because the target has not been reached, record that as a decision rather than presenting it as new predictive evidence.
The purpose of this guide is to assess a forecasting process, not to rank employees or automate decisions about their pay or rights. Discuss access and appropriate use with the people responsible for those policies. Commercial estimates can contain personal and customer information, so a broader audience does not automatically need every underlying record.
Make uncertainty usable rather than decorative
Ask the person receiving the pack to explain what would make them choose a different action. If they cannot connect the range or scenario to the decision, simplify the presentation or revisit the brief. An extra shaded band does not help merely because it makes a chart look more analytical.
For a reporting implementation, HubSpot reporting is relevant to keeping the distinct views and their definitions visible. Require a reader to identify the official estimate, its alternatives and the original assumptions without opening a hidden worksheet. That is a practical test of whether the uncertainty display communicates anything useful.
Check what your CRM actually records
Static graphic. Enable JavaScript to explore the stages.
A platform can provide a forecast screen, submission history and an accuracy measure without automatically answering your exact buying question. Inspect the definitions and saved records behind those features. The important distinction is between what the product records natively and what your fixed-cutoff evaluation requires it to preserve or export.
Treat a demonstration as a chance to trace one result, not a tour of every dashboard. Bring a small test case with known dates and amounts. Ask the supplier to show where each value appears, when it was recorded and how it contributes to the displayed forecast and later score.
Read the native score definition
As checked in November 2026, HubSpot’s forecast accuracy documentation describes a submission average for the selected forecast period and a closed-won comparison. That average is not necessarily the particular submission saved at your chosen decision cutoff. Confirm the selected period, pipeline, history and access available in your account before treating the native score as your acceptance measure.
For a hypothetical illustration, suppose a user submits £100,000 early in a period and £160,000 later, with a final outcome of £130,000. The average of those two submissions is £130,000. That average matches the outcome, but the early fixed-date forecast was £30,000 low and the later submission was £30,000 high.
Neither view needs to be wrong. They describe different questions. The average-submission view summarises the submitted numbers for that period; the fixed-date test judges the estimate available when a particular decision was made. Label the views clearly and avoid using one to imply that the other has been demonstrated.
Do not recreate a vendor’s displayed score from memory. Read its current definition and compare a known example in the actual product. If the documentation leaves an edge case unclear, record the question and test it in the permitted environment. A zero-actual period, for example, should not be given an invented handling rule in your specification.
Ask for the evidence to leave the dashboard
Request an export that includes the identifiers needed to reproduce the comparison. The required fields follow from your forecast brief and snapshot design, not from whichever columns happen to appear by default. A screenshot of a score is useful presentation evidence, but it is not a substitute for the records that produced it.
Check the relationship between submissions, users, teams and periods. If a hierarchy changes, can the historical view still explain which structure applied at the time? Decide whether historical reporting follows the original structure or a restated current structure, and label the choice. Mixing the two can make period comparisons difficult to interpret.
Microsoft’s Dynamics 365 forecast configuration guidance describes forecast hierarchy, access permissions and grid columns as configuration choices. It also warns against using the feature for employment decisions affecting compensation or other rights. These details underline the need to inspect both the report’s meaning and its intended use, rather than assuming every available view suits every decision.
Ask how exports remain available if a supplier leaves or an individual account changes. Confirm ownership, authorised access and the process for recovering the calculation pack. Do not ask staff to share credentials to make a demonstration work. The business needs an approved operating route that survives ordinary personnel and supplier changes.
Demonstrate awkward records alongside clean ones
Give the supplier a deal whose expected date changes after the cutoff. Retrieve the historical snapshot and the current record side by side. The snapshot should show the admitted historical value, while the current record should show the later state. The two views need labels that prevent a reviewer from confusing them.
Next, provide a missing or late input. Ask where the exception appears, whether the forecast still runs and how the operator knows which information was unavailable. A system that completes successfully while silently dropping material inputs has not explained its output merely because the calculation returned a number.
Test a record that is excluded by the agreed population rule. The exclusion should be reproducible and visible to an authorised reviewer. Then change the controlled test so the record qualifies. You should be able to explain the change in membership without editing a hidden list or asking the original developer to intervene.
Use a controlled duplicate or correction case as well. Confirm how it affects the current actuals view and whether the previously released actual remains traceable. The aim is to understand the product’s behaviour, not to introduce unsafe test records into a live sales process. Agree the test environment and avoid sending customer communications during a demonstration.
Check access at the level of the evidence
A user may be allowed to see an aggregate forecast without needing every underlying customer record. Demonstrate the intended boundary with representative roles. Check the report, the export and any linked record view, because a sensible-looking dashboard does not prove that its supporting files have the same restriction.
Keep administrative configuration separate from ordinary use where the platform supports that distinction. Identify who can change definitions, edit forecast settings and view historical evidence. Those permissions affect whether the agreed process can be maintained consistently. Review the actual account setup instead of relying on a generic product brochure.
Licensing and feature eligibility should be confirmed for the account and proposed users when purchasing. Do not assume that a feature shown in a demonstration is included in every plan or seat. Ask for the relevant scope in the supplier’s written proposal, without treating this guide as a current price list.
If the wider account needs implementation work, the HubSpot implementation guide covers that broader buying task. Here the acceptance question remains specific: can your approved users retrieve the right dated evidence and reproduce the comparison the business actually needs? A platform badge or a polished forecast screen does not answer that by itself.
Commission a forecast you can accept and maintain
Static graphic. Enable JavaScript to explore the stages.
A good proposal should explain what will be delivered, how you will test it and what happens after the first release. Predictive usefulness cannot always be proved at installation, especially when clean historical snapshots do not exist. Separate the work of implementing the agreed process from the later evidence needed to decide how widely to rely on it.
Make that distinction visible in the acceptance plan. Functional tests can show that snapshots, calculations, access and exports behave as specified. Predictive evaluation asks how the dated estimates compare with later outcomes and a baseline. A supplier should not use a successful installation test as proof of future accuracy.
Request a buyer-owned evidence pack
The pack should make a forecast understandable without the original consultant narrating it. Include the decision brief, data definitions, input availability rules, method versions, sample snapshots and calculations. Keep the agreed exclusions and known limitations beside the results rather than in an appendix nobody can connect to the displayed score.
Ask for an inventory of the delivered components and their owners. Identify the location of the snapshot store, the authorised export process, the calculation configuration and the operational instructions. The business should know which components it controls and which depend on a supplier or platform service.
Require an example that another authorised person can reproduce. Give that reviewer the saved pack and the written instructions, then ask them to calculate a selected forecast and score. Record any missing step. This tests the quality of the handover more directly than asking whether the documentation looks comprehensive.
The proposal should also explain what is not included. Historical reconstruction, data cleansing, new integrations and ongoing model changes can be separate pieces of work. Ask where the initial scope stops and which gaps would prevent evaluation. That avoids treating every later discovery as either an unplanned extra or an unexplained defect.
Define acceptance in two parts
Use clear functional pass conditions for the process. A saved snapshot retains its original values after a current record changes. A late input follows the stated availability rule. An authorised operator can export the agreed evidence. A calculation can be reproduced within the stated tolerance. These conditions can be tested without inventing an industry-wide accuracy target.
Set the predictive criteria separately, before the final comparison. Name the outcomes, horizons, evaluation periods, preferred measures and baseline. Explain how incomplete forecasts and changed populations will be handled. If the evidence will be collected prospectively, state what decision will be made when the agreed observations become available.
Do not turn a threshold into a guarantee. A method can meet the agreed criteria on the evaluation set without being certain to do so in every future period. The acceptance record should describe the demonstrated evidence and the intended use, including any restrictions, rather than promising that the model has eliminated commercial uncertainty.
A useful acceptance checklist can include these specific demonstrations:
- Reconstruct a named historical forecast using only its admitted inputs and saved version.
- Show that a later date or value change does not overwrite the original snapshot.
- Recalculate a score from the exported forecast and actual records.
- Explain one excluded record and one incomplete or failed forecast.
- Demonstrate the intended access boundaries and an approved operating fallback.
Plan the first operating review
Choose a review schedule that fits the forecast cycle and the availability of agreed actuals. There is no benefit in holding a formal accuracy review before the outcome definition allows the relevant periods to be complete. Leaving known errors unexamined until a distant renewal meeting can make the operating evidence harder to recover.
Give the review an owner and a short agenda. Check whether scheduled snapshots were produced, whether inputs were missing, which actuals changed and how the latest completed forecasts compared with the baseline. Separate operational incidents from predictive misses so each receives an appropriate response.
Keep actions attached to evidence. If a missed booking appears to result from a late source update, identify the relevant timestamps and records before changing the forecasting method. If the calculation used the wrong amount basis, fix that definition and assess which previous results are affected. Different causes need different follow-up tests.
Retain a record of the review decision: continue, investigate, restrict a use, revert to a documented fallback or evaluate a revised version. This creates an operating history that a new owner can understand. A list of scores without the decisions made from them is a thinner handover than it first appears.
Control changes after launch
A new field, altered sales process or supplier model update can change the forecasting task. Record the effective date and the affected components. Ask whether the existing evaluation evidence still applies, or whether a new bounded test is needed before the revised method is used for the same decisions.
Keep the previous working version and its evidence retrievable through the organisation’s approved arrangements. Reverting a calculation is different from restoring old source data, and both are different from changing a past published forecast. The operating instructions should make those distinctions explicit so an incident response does not destroy the historical record.
For a proposed revision, compare its results with the existing version under the same information boundary. Preserve the original acceptance criteria or document why they have changed. A new model name or release number should not allow the evaluation to restart with a more convenient population and no explanation.
Include an exit exercise before the project is treated as complete. Ask an authorised internal operator to locate a snapshot, run an export, explain a score and identify the support route for a failure. If those tasks still require the supplier to operate an undocumented tool, the handover has a practical dependency that belongs in the acceptance decision.
The CRM data cleansing guide can help frame source-quality work identified during that process. Keep any such work distinct from proof of predictive improvement. A maintained forecast is one whose definitions, evidence and operating responsibilities remain clear when the business changes, not merely one that produced an attractive result at launch.
Questions buyers ask about sales forecasting
Static graphic. Enable JavaScript to explore the stages.
What is a useful accuracy target for our business?
Start with the decision, its lead time and the cost of the mistakes that matter. A money error that is manageable for a combined quarterly view may be unhelpful for a monthly service-line decision. Agree the measure, population and baseline before setting an acceptance threshold.
Ask a supplier to explain its proposed target using your evidence, rather than adopting an unsupported industry percentage. Keep functional requirements separate: a forecast should still be reconstructable and delivered on time even when its predictive performance is uncertain. If the available evaluation periods are too limited for a firm conclusion, narrow the intended use or collect more prospective evidence.
Can we start without historical forecast snapshots?
Yes, but be clear about what the initial work can prove. Reliable completed outcomes may support a simple historical reference. They do not automatically reveal the opportunity values, assumptions or customer information available at earlier cutoffs. A reconstructed pipeline test needs those limitations stated.
Begin saving dated forecasts and their admitted inputs while the existing process continues. Compare the candidate and baseline prospectively once the corresponding outcomes are agreed. This creates evidence that does not depend on reconstructing forgotten decisions. The observation period should follow your forecast horizon and commercial cycle, not an arbitrary promise that every implementation will prove its value within the same number of weeks.
Can HubSpot report accuracy at our chosen fixed date?
Check what your account actually preserves and exports. A native score based on an average of submissions is a different measure from the error of one submission saved at a named cutoff. Do not assume that a historical score answers the fixed-date question simply because it concerns the same month.
Ask for a demonstration using a known period with more than one submission. Identify the original cutoff value, the later changes and the actuals definition, then reproduce the comparison outside the dashboard. If the needed historical evidence is unavailable, record that gap and begin collecting it prospectively rather than presenting an inferred value as an original submission.
How do we compare an AI forecast with a simple one?
Give both methods the same outcome definition, information cutoff and target horizon. Preserve their predictions before the result is known, then compare the completed cases using the measures agreed in advance. Include unavailable or late forecasts in the operating record instead of showing only successful calculations.
The AI label does not remove the need for a baseline, a readable explanation of inputs or an exportable evaluation pack. Ask what extra maintenance and dependencies the method introduces, and what evidence would justify them. If the supplier changes its model during the test, retain the version history and explain which results belong to each version before combining them.
What should the supplier hand over at completion?
Request enough evidence for an authorised colleague to reconstruct a forecast, explain its score and operate the next cycle. That includes definitions, input availability rules, method versions, example snapshots, calculation instructions, known limitations and the agreed access arrangements. The pack should also identify ownership and the support route for failures.
Demonstrate the handover with a real acceptance exercise rather than relying only on a document checklist. Ask someone who did not build the system to retrieve a historical record and reproduce its result. Record any dependency on the supplier, including exports or configuration that the business cannot access through its approved operating arrangements.
Choose the next step with a forecast you can check
Static graphic. Enable JavaScript to explore the stages.
A useful forecast leaves an evidence trail you can inspect. Begin with the decision that needs an earlier view, then choose a bounded comparison that your available records can support. A sophisticated method is worth considering when its evidence and operating demands fit that task.
- Define the outcome, amount basis, cutoff and target horizon before selecting a method.
- Preserve original snapshots and distinguish later revisions from the evidence available then.
- Keep a simple baseline and compare methods on the same dated cases.
- Inspect error direction, magnitude and relevant breakdowns rather than one headline score.
- Require a reproducible handover, named ownership and a documented fallback.
If you want to clarify those requirements for your CRM and sales process, book a GRO strategy meeting to discuss the forecast and comparison work your business needs.