Forecast accuracy compares each period's forecast against the actual result and reports the gap as an error rate, most often mean absolute percentage error. The number only means something if the forecast was locked before the period started and the actuals were closed on a consistent basis.
How forecast accuracy is measured
The method is the same everywhere. Lock the forecast for a period, wait for the period to close, then compare the forecast against the actual and express the difference as a percentage. Aggregate those period errors and you have an accuracy rate. The most common measure is mean absolute percentage error, and the definitions of MAPE alongside the measures built to correct for its weaknesses are set out in the open-access forecasting text Forecasting: Principles and Practice. None of that is contested and none of it is where companies go wrong.
The two conditions that actually matter are unglamorous. The forecast has to be locked before the period starts, because a forecast revised mid-period grades itself. And the actuals it is compared against have to be closed and reconciled on a consistent basis, which is the recognition question the AASB accounting standards govern. If the close moves revenue between months, your accuracy number is measuring your close, not your forecast.
Accuracy is only as good as the actuals behind it, which is why who owns the profit and loss statement decides whether the comparison means anything.
Over-forecasting and under-forecasting say different things
An absolute error rate hides the direction, and the direction is the useful part. Persistent over-forecasting on revenue usually means the assumptions were set to support a story rather than a plan. Persistent under-forecasting on costs is more dangerous, because it flatters runway and the business only finds out late.
Both are visible in the same place. What investors interrogate in a founder built model is not the output but the ramp: flat-phasing every cost line looks amateur, because a cost base scales as you hire rather than arriving on day one.[1] A model that phases costs flat will under-forecast spend every single month, and the error rate will be the last thing to tell you.
| What the error pattern looks like | What it usually means | |
|---|---|---|
| Revenue forecast consistently high | Over-forecasting, month after month | The assumptions were built to support a target rather than derived from the business. Check whether the growth rate traces back to anything real. |
| Cost forecast consistently low | Under-forecasting, month after month | Usually flat-phased costs or an uncosted hiring plan. This one flatters runway, so it is found late and hurts most. |
| Error swings either way | Large errors with no consistent direction | The model is not the problem. Either the actuals are moving between periods at close, or the forecast is being revised inside the period it is meant to be graded against. |
| Accurate in aggregate, wrong in the peaks | A good annual number hiding bad months | Seasonality is not modelled. Report accuracy by period, because the months you get wrong are usually the months that decide the year. |
What accuracy is realistic at each horizon
Accuracy is a function of distance. In an early-stage business the next twelve months should be held to a high standard because you have conviction on them, months twelve to twenty-four to line of sight, and anything past that is a bonus rather than a target.[2]Grading a month thirty-six forecast against the actual is theatre. Nobody learns anything from it and it teaches the team to hedge.
Report it by period as well as in aggregate, because a single blended number smooths over exactly the seasonality you needed to see. A business that is accurate for nine months and badly wrong in the three that matter has a worse problem than its annual figure suggests.
If the horizon problem keeps repeating, the usual fix is moving from an annual budget to a rolling cost forecast.
Who owns forecast accuracy in a growing company
In most companies at this stage the honest answer is the founder, and that is the problem. Founder-led finance does not scale at all. It runs its course as the business grows, and what it leaves behind is a scrambled financial model and cash conversations that are closer to a guess than a position.[3] No metric fixes that. A person does.
The pattern shows up most sharply straight after a raise. The forecast is suddenly not detailed enough for the new board, the reporting rhythm is still calibrated to a business half the current size, the hiring plan has not been properly costed, and the cash discussion is running at the level it was twelve months ago.[4] That is not a modelling failure. It is a signal that the work has outgrown whoever is currently doing it.
What this means for who you hire
Accuracy is a proxy for judgement, so it is worth treating as a hiring signal rather than a dashboard tile. When candidates put numbers to achievements on a CV, showing forecast accuracy improved from 50% to 95% is a genuinely good example, because it demonstrates the level of impact and value they brought to the role.[5] Ask for it in interviews and ask what they changed to get it.
The same test applies to the tooling. In finance, AI handles the basic work well, checking, reconciliation, producing a first-pass analysis or forecast, but the final stage still needs a qualified human who understands the challenge and what a good outcome looks like.[6] An automated forecast raises the floor on accuracy. It does not remove the need for someone who can defend the assumptions when the board asks where the number came from.
The practical version of this question is when to make your first senior finance hire, and the forecast is usually the thing that forces it.
Common questions
How is forecast accuracy calculated?
Lock the forecast before the period starts, wait for the period to close, then compare the forecast against the reconciled actual and express the gap as a percentage. Aggregating those period errors gives an accuracy rate, most commonly reported as mean absolute percentage error. Two conditions decide whether the number means anything: the forecast must not be revised inside the period it is graded against, and the actuals must be prepared on a consistent basis.
What is a good forecast accuracy percentage?
It depends entirely on the horizon and the business. In an early-stage company the next twelve months should be held to a high standard because you have conviction on them, months twelve to twenty-four to line of sight, and anything beyond that is a bonus rather than a target. Chasing a single headline accuracy figure across a three-year model teaches a team to hedge rather than to think.
Why does our forecast keep missing after a funding round?
Because the round changed the audience and the pace, not the finance function. The common pattern straight after a raise is a forecast that is no longer detailed enough for the new board, a reporting rhythm calibrated to a business half the current size, a hiring plan that was never properly costed, and cash conversations still running at last year's level. That is a resourcing signal rather than a modelling one.
Can AI improve forecast accuracy?
It improves the floor. AI handles checking, reconciliation and producing a first-pass analysis or forecast well, which removes a lot of the mechanical error. Defending an assumption to a board is the part it cannot do, so the final stage still needs a qualified human who understands the challenge and what a good outcome looks like.
References
- Daniel Ross, advisor at Triple Bubble and Euphemia, on cost ramp and deal terms, in Story Recruitment's panel on what investors look for in founder built financial models.
- Daniel Ross on forecast horizon: twelve months of high conviction, twelve to twenty-four months of line of sight, and accuracy beyond that as a bonus. From the same Story Recruitment panel on founder built financial models.
- My position on it: it does not scale at all, and eventually runs its course as a business grows, leaving a scrambled financial model and cash conversations that are more of a guess.
- A pattern Tom Hunter sees after a funding round: the forecast is not detailed enough for the new board, the reporting rhythm is still calibrated to a business half its current size, an ambitious hiring plan has not been properly costed, and cash discussions remain at the same level as twelve months prior.
- The advice I give on this: it is great to put a few more metrics to achievements on a CV, such as showing forecast accuracy improved from 50% to 95%, to illustrate the level of impact and value a candidate brought to the role.
- Tom Hunter on AI in finance: it can easily handle basic tasks like checking, reconciliation, or producing basic level analysis, forecasts or models, but you still need a qualified human in the loop for the final stages who understands the challenge and what a good outcome looks like. From his appearance on Celia's Corner.