When Was Your Last Forecast Correct?

Ask a leadership team when their last forecast was correct and the room usually splits. Finance remembers the quarter it landed. Sales remembers the quarter it missed. Operations says the number changed four times, so which one?

That disagreement is the real answer. Most companies cannot say when a forecast was last right because nobody wrote down what "right" meant before the period started.

They are not alone in doubting the number. In Gartner's State of Sales Operations Survey, only 45% of sales leaders and sellers said they had high confidence in their organization's forecasting accuracy.^[1] Gartner warned that this level of skepticism can lead to decisions based on intuition instead of evidence.^[1]

Define "correct" before the period starts

A forecast cannot be graded after the fact if the test was never set. Before the quarter opens, agree on four things:

• The number. Bookings, revenue, units, renewals. One definition, from one system.
• The level. Company total, region, product, or rep. A forecast can be right in total and wrong everywhere underneath.
• The horizon. The forecast made 90 days out and the one made the week before are different forecasts. Grade each one separately.
• The tolerance. How far off can it be before a decision goes wrong? That answer comes from the business, not the data team.

The measure matters too. The forecasting textbook by Rob Hyndman and George Athanasopoulos points out that percentage error, one of the most common scores, becomes infinite or undefined when an actual value is zero and gives extreme values when actuals are close to zero.^[3] If you forecast new product lines or small accounts, a single month near zero can make a sound forecast look like a disaster. Pick the score with your data team and keep it fixed so the trend means something.

Keep every version of the forecast

Most companies keep only the final forecast. That hides the most useful information.

A forecast usually passes through several hands: a system or model produces a baseline, an analyst adjusts it, sales adds a view, and leadership signs off. Michael Gilliland of SAS calls the method for grading those steps forecast value added: the change in accuracy that can be attributed to each step and each participant in the process.^[4] To do it, you keep the forecast at every step alongside what actually happened.^[4]

The results can be uncomfortable. The SAS paper describes a specialty retailer whose analyst overrides beat a simple five-week moving average only 25 percent of the time.^[4] It describes a food manufacturer that asked salespeople for item-level forecasts on its biggest customers, found after several months that their input had no impact on accuracy, and went back to its original process.^[4] The paper's own summary: it is quite common to find that executive tampering with a forecast just makes it worse.^[4]

That does not mean judgment has no place. A study by Robert Fildes, Paul Goodwin, Michael Lawrence, and Konstantinos Nikolopoulos collected more than 60,000 forecasts and outcomes from four supply chain companies.^[5] In three of the four, manual adjustments improved accuracy on average.^[5] But small adjustments often damaged accuracy, and upward adjustments were much less likely to help than downward ones and were more often made in the wrong direction, which the authors read as a general bias toward optimism.^[5]

A forecast that gets touched by five people should be graded five times.

The practical lesson: let people adjust when they know something the model cannot, like a signed contract or a lost distributor. Record the reason. Then grade the adjustment, not just the final number.

Beat the simple forecast or stop paying for it

Every forecast needs a baseline that costs nothing. The simplest is "next period looks like this period." For seasonal businesses, it is "this month looks like the same month last year."

Hyndman and Athanasopoulos say the plain version works remarkably well for many economic and financial series, and they use these simple methods as benchmarks: if a new method cannot beat them, it is not worth considering.^[3] The SAS paper reports that a large technology manufacturer reviewed six years of performance and found that half the time a simple forecast did as well as or better than its official forecast.^[4] It also cites a 2014 study by Steve Morlidge of eight companies and more than 300,000 forecasts, in which 52 percent failed to beat the simple forecast.^[4]

Gilliland puts the takeaway in one line: if you don't know that you are beating a naïve forecast, then maybe you're not.^[4]

This is the cheapest test in forecasting. Rebuild the simple forecast for last year, which takes an afternoon, and compare it with what your team actually published. The SAS paper suggests a full year of data before drawing firm conclusions, since short windows bounce around on randomness.^[4]

Grade it on months it never saw

A forecast that explains the past is not the same as one that predicts the future. Hyndman and Athanasopoulos are direct about it: a model that fits the training data well will not necessarily forecast well, and accuracy can only be judged on new data that was not used to build the model.^[3]

Ask how the accuracy you were shown was measured. If the answer is "on the same history the model learned from," you have not seen its accuracy yet. (The method we use, testing on rolling periods the model never saw, is in Why Forecasting Models Degrade.)

Own the model, not just the number

Many companies receive a forecast they do not control. It comes out of a vendor tool, a spreadsheet one analyst built, or a CRM rollup of rep commits. When it misses, nobody can say why, and nobody can change it.

Owning the model does not mean building everything in house. It means your company can answer a few questions on demand:

• What data is it trained on, and through what date?
• Who is allowed to change it, and when was it last changed?
• What did it predict for each of the last eight periods, and what happened?
• Can we retrain it this month if the market shifts?

If those answers live with a vendor or a former employee, you are renting the number. (Why the training date matters: When Were Your Models Last Updated?.)

Align the data science with the decision

A forecast exists to make a decision better. Start there and work backward.

• Hiring and capacity need a forecast at the team level, far enough out to recruit and ramp.
• Inventory and staffing need it at the item or location level, a few weeks out.
• Cash planning needs a conservative floor more than a best guess.

The cost of missing high is rarely the same as the cost of missing low. The SAS paper describes an automotive supplier that weighed both sides: forecasts that run high bring excess inventory, and forecasts that run low bring unfilled orders, lost revenue, and lost credibility.^[4] When the costs are uneven, tell your data team which side hurts more. That one sentence changes the model they should build.

Then show the track record. Gartner advises sales operations leaders to go beyond publishing regular forecasts and also track and communicate accuracy over time.^[2] Trust in a forecast is earned with a visible history, not a confident presentation.

What to ask for

1. A written definition of "correct" for each forecast: the number, the level, the horizon, and the tolerance.
2. A fixed accuracy score chosen with your data team and kept the same quarter to quarter.
3. Every version saved: the baseline, each adjustment, the final number, and the actual.
4. A simple baseline rebuilt for last year and compared with what you published.
5. Accuracy measured on unseen periods, not the history the model learned from.
6. A named owner who can explain, retrain, and change the model.
7. A one-page track record of forecast against actual, shared with leadership every period.

Then ask the question again. When was your last forecast correct? You should be able to answer with a date.

If you want forecasts your leadership team can check instead of argue about, talk to Gamify Data. We start with the data you already have and build a forecast your team owns.

Sources

1. Gartner, "Gartner Says Less Than 50% of Sales Leaders and Sellers Have High Confidence in Forecasting Accuracy," press release, February 12, 2020. https://www.gartner.com/en/newsroom/press-releases/2020-02-12-gartner-says-less-than-50--of-sales-leaders-and-selle
2. Jordan Bryan, "5 Ways to Develop a World-Class Sales Operations Function," Gartner, November 18, 2020. https://www.gartner.com/smarterwithgartner/5-ways-to-develop-a-world-class-sales-operations-function
3. Rob J. Hyndman and George Athanasopoulos, "The forecaster's toolbox," chapter 5 of Forecasting: Principles and Practice, 3rd edition, OTexts, accessed October 26, 2026. https://otexts.com/fpp3/accuracy.html
4. Michael Gilliland, "Forecast Value Added Analysis: Step by Step," SAS Institute white paper, 2015. https://www.sas.com/content/dam/SAS/en_us/doc/whitepaper1/forecast-value-added-analysis-106186.pdf
5. Robert Fildes, Paul Goodwin, Michael Lawrence, and Konstantinos Nikolopoulos, "Effective forecasting and judgmental adjustments: an empirical evaluation and strategies for improvement in supply-chain planning," International Journal of Forecasting, vol. 25, no. 1, 2009, pp. 3 to 23. https://doi.org/10.1016/j.ijforecast.2008.11.010