Forecast Accuracy
In hotel revenue management, forecast accuracy is how close your predictions came to what actually happened, measured systematically against the forecasts you wrote down rather than against the ones you remember making.
The useful finding is almost never how far off you were. It is that you are off in the same direction every time, on the same kind of date, which is a fixable problem hiding inside an unfixable one.
See how ampliphi prices for independent hotels
Get a free revenue audit · Book a demo
What is forecast accuracy?
The gap between forecast and outcome, tracked over enough dates to mean something.
It requires a thing most properties do not have: forecasts recorded with the date they were made. A forecast you did not write down teaches you nothing afterwards, because memory reliably adjusts to whatever happened. You will remember having expected it.
Two measures matter and they answer different questions. Error is how far off you were, usually reported as a mean absolute difference in rooms or occupancy points, and it tells you how much confidence the forecast deserves. Bias is whether your errors point the same way, measured as a signed mean, and it tells you whether something in your method is systematically wrong.
That distinction is the whole reason to do this. Error at a small property is largely irreducible, because one group booking is several occupancy points and no method removes that. Bias is a defect. If you forecast midweek dates four rooms low every time, that is not noise, it is a correction waiting to be applied.
Accuracy also has to be stated at a horizon. A forecast made 90 days out and one made 7 days out are different claims, and comparing them tells you nothing except that the second is easier. Track each horizon separately.
Resources: Demand forecasting · On the books · Booking pace
How forecast accuracy works in practice
You log the forecast, log the outcome, and look at the signed difference rather than the absolute one.
Worked example. A 46-room inn logs its 30-day-out forecasts for a year. Mean absolute error comes out at 4.2 rooms, which looks poor against a 46-room house. The signed mean is minus 2.8, meaning it forecasts low more often than high. Split by date type, weekends run at minus 0.4 and midweek at minus 3.9. The property is systematically pessimistic about midweek dates, which explains a long-standing habit of discounting Tuesdays and Wednesdays that would have filled anyway. Correcting that bias is a single adjustment. Chasing the 4.2 is a project with no end.
The split is where the value appeared, and it would have been invisible in the headline number.
The everyday practice is three columns and a habit. Date, forecast, date the forecast was made. Then after the date passes, the actual. Thirty seconds a week, and after a season you have something worth reading.
What you look for is patterns rather than misses. Individual dates that came in wrong are noise. A whole category of dates that comes in wrong the same way is a method problem, and the categories worth splitting by are day of week, season, lead time and whether the date had a known event.
The second habit is being honest about contaminated forecasts. If you forecast a date at 30, then discounted it and it finished at 38, your forecast was not wrong in an interesting way. You changed the thing you were predicting. Mark those rather than letting them flatter or damage the record.
Resources: Need date · Rate calendar
Why forecast accuracy matters for independent hotels
Because the alternative is a pricing instinct that never gets corrected.
An owner-operator forecasts constantly, mostly informally, every time they look at a date and decide whether it needs help. Those judgments are the pricing process at most small properties. If they are consistently pessimistic about midweek, the property will discount midweek for years, and nothing in any report will ever say so.
Measuring is what turns that from an instinct into something that improves. It is also unusually cheap: no software, no analysis, a spreadsheet with four columns and a few seconds a week.
The other reason is that it tells you how much to trust the forecast when it matters. A displacement decision resolved by five rooms is meaningless if your 30-day error is routinely four. Knowing your own error bar is what stops you treating a marginal calculation as a verdict.
The expectation should be modest. At 46 rooms a mean absolute error of three or four rooms at 30 days out is respectable, and precision beyond that is not available at this scale. The goal is an unbiased forecast with known noise, not an accurate one.
Resources: Revenue management for independent hotels
How to measure forecast accuracy at your property
- Write every forecast down with the date it was made. Nothing else here works without this.
- Track by horizon. A 30-day forecast and a 7-day forecast are different claims and should not be pooled.
- Use the signed difference, not the absolute one. Absolute error hides bias, and bias is the fixable part.
- Split by day of week, season and lead time. The headline number almost never contains the finding.
- Mark forecasts you then acted on. A date you discounted is not a clean test of the prediction.
- Correct bias, accept noise. A consistent direction is a defect. Individual misses at 46 rooms are arithmetic.
What forecast accuracy will not tell you
It does not tell you why a forecast missed. A date that came in ten rooms under could be a competitor opening, an event that moved, weather or nothing at all, and the error figure contains none of it.
It cannot be compared across properties. Error scales with room count, with how volatile a market is and with how much group business a property carries, so a benchmark from elsewhere is close to meaningless. Compare only against your own history.
And accuracy is not the objective. A more accurate forecast that changes no decision was not worth the effort, and a rough forecast that stopped you discounting a date unnecessarily paid for itself. The measure is better decisions, not a smaller number.
How ampliphi approaches forecast accuracy
Ampliphi does not publish an occupancy forecast, so there is no forecast accuracy figure for it to report. That follows directly from what the product does: the everyday suggestion is a rate, built on booking pace and occupancy, rather than a prediction of how a date will finish.
The distinction matters when comparing products. An enterprise revenue management system that forecasts demand by segment can be evaluated on forecast error, and asking for that figure is a reasonable question to put to one. Asking it of a system that does not forecast is a category error, and a vendor answering it anyway is worth a second question.
What this means for you is that the forecasting habit above stays useful regardless. Your own written forecasts, and the bias you find in them, inform staffing, ordering and group decisions that no rate suggestion addresses. The suggestion covers your base rate and the differential between room types, you approve every rate before it publishes, and ampliphi runs on top of the PMS you already use.
Key takeaways: forecast accuracy
- How close your forecasts came to reality, measured against ones you wrote down.
- Two measures: error, meaning how far off, and bias, meaning consistently off in one direction.
- Bias is a defect and is fixable. Error at 46 rooms is largely arithmetic and is not.
- Use the signed difference. Absolute error hides the finding.
- Split by day of week, season and lead time. The headline figure rarely contains anything useful.
- Knowing your own error bar is what stops you treating a marginal calculation as a verdict.
Frequently asked questions about forecast accuracy
How accurate should a hotel forecast be?
Accurate enough to change decisions correctly, which at a small property is a lower bar than it sounds.
At 46 rooms, a mean absolute error of three or four rooms at 30 days out is a reasonable expectation. One group booking is eight rooms, so the irreducible noise is large relative to the house, and no method removes it.
Larger properties achieve tighter figures because their volumes are steadier, not because their methods are better in a way that transfers. Judge yourself against your own history rather than against a published benchmark from properties that do not resemble yours.
What is the difference between forecast error and forecast bias?
Error is magnitude. Bias is direction.
Mean absolute error tells you the typical size of a miss, which is how much confidence the forecast deserves. Signed mean error tells you whether the misses cancel out or pile up on one side, which is whether something in your method is systematically wrong.
A property with a four-room error and no bias is making noisy but honest forecasts. A property with a four-room error and a minus three bias is understating demand nearly every time, and correcting that is a single adjustment rather than a research project.
How do I track forecast accuracy without software?
Four columns in a spreadsheet: the date, the forecast, the date you made the forecast, and eventually the actual.
Add the row when you make the forecast and fill in the actual after the date passes. That is a few seconds a week. After a season you can calculate a signed mean and split it by day of week, which is where the useful finding almost always is.
The discipline is adding the row at the time. A forecast reconstructed later is not a forecast, because you already know how the date turned out.
Does measuring forecast accuracy actually improve anything?
It improves bias, which is usually where the money is.
Finding that you forecast midweek dates four rooms low every time explains a years-long habit of unnecessary discounting, and correcting it costs nothing. That kind of finding is common at properties measuring for the first time and impossible to see without a record.
It will not reduce your noise much. At 46 rooms, individual misses are mostly the arithmetic of a small house, and chasing them produces effort without improvement. Fix the direction and accept the spread.
What if I acted on a forecast and changed the outcome?
Mark it and treat it separately, because it is no longer a test of the prediction.
A date forecast at 30 that you then discounted, promoted or restricted, finishing at 38, tells you something about your intervention and nothing clean about your forecasting. Pooling those with untouched dates will make your forecasts look worse than they are or better, depending on which way you intervened.
Keep a flag column. The untouched dates are your accuracy measure. The touched ones are a record of whether your interventions worked, which is a separate and equally worthwhile thing to know.
Related terms
Demand forecasting
Predicting how a future date will finish. Accuracy is the practice of checking afterwards whether it worked.
Booking pace
How a date is filling against its own history. The main input to a forecast, and where a systematic bias usually originates.
On the books
The count of rooms reserved for a future date. The starting term in the forecast, and an error here propagates into everything downstream.
Occupancy rate
The percentage of available rooms sold. The outcome your forecast is measured against.
Need date
A date forecast to underperform. Forecast bias is why properties find need dates that were never actually soft.