Forecast Accuracy tells you how close your past forecasts came to what actually sold. Every week Toolio saves the forecast it served, then scores it once the sales are in. Use it to see how well a forecast held up over the weeks after it was made, which channels forecast well, and whether you are consistently over or under.

Unlike Forecast Confidence Score, which estimates how much to trust a method before you plan, Forecast Accuracy reports what already happened. The two answer different questions and are best read together.

Every week Toolio saves the forecast it is serving for each choice and cluster. That saved copy is a snapshot, dated by the week it was taken.

The snapshot holds the System forecast: the forecast the AI or Statistical engine produced. Changes a planner makes to the forecast in the plan are not part of the snapshot, so Forecast Accuracy measures how well the engine forecast, not how well the edited plan did. Sales are compared as gross sales units.

A forecast window is a span of weeks that follows a snapshot. A forecast window of four weeks asks one question: how did the forecast we were looking at four weeks ago hold up over the four weeks since? So the four-week score reads the snapshot from four weeks ago and compares it against the last four weeks of sales. The 13-week score reads the snapshot from a quarter ago and compares it against the whole quarter that followed it.

That is why accuracy is reported per forecast window rather than as one blended number: each window grades an older snapshot over a longer span.

A forecast window is not a lead time. The 13-week score is not the accuracy of the forecast made 13 weeks before each selling week. It grades one snapshot over the quarter that followed, so it mixes forecasts made 1 to 13 weeks ahead. See the next section.

Scores are never back-filled or approximated. A forecast window is scored only when a snapshot exists for exactly that week — a snapshot one week off is not substituted, because the number would still look plausible while meaning something else. Accuracy therefore accrues from the week snapshotting started on your account.

You may know forecast lag from other tools. A lag-4 score compares a single week: the week that was four weeks out when the forecast was made. Toolio does not score lags. Each forecast window is a horizon: it scores every week from the snapshot up to the end of the window.

So a 13-week score tells you how well the forecast from a quarter ago held up across the whole quarter, not only at its far end. The last four weeks of that quarter are the same weeks the 4-week score measures, read from an older snapshot.

This is also different from the Forecast Horizon in Assortment Plan Settings (Max Forecast Length). That setting controls how far ahead Toolio forecasts at all. A forecast window is the span of weeks a snapshot is scored over.

Every scored cell shows two numbers, and they mean different things. Both are measured week by week, never on the totals alone.

Accuracy is how much the forecast and actual sales overlapped each week: the smaller of the two summed across the weeks, over the larger of the two summed across the weeks.

100% means every week matched exactly. A 2× overforecast scores 50%, a 3× overforecast 33%.

Bias is the direction of the error — total forecast against total sales.

Positive means you forecast too high, negative too low.

Read them together, because they can disagree and the disagreement is the useful part. Take a four-week window where you forecast 100 units a week:

Bias is 0% — the total was exactly right. Accuracy is 74%, because the overlap is 340 units out of a possible 460. The level was right and the shape was wrong, which is a phasing problem rather than a demand problem. Bias alone would have called this a perfect forecast.

Accuracy is shown at three forecast windows: 4, 8 and 13 weeks.

They fill in order. Four weeks is available first, and each longer forecast window becomes available that many weeks after snapshotting began — so the 13-week column completes a quarter in. Empty long columns are expected on a new account, not a fault.

Toolio also scores two longer forecast windows, 26 and 52 weeks, but hides them by default. A 26-week score needs a snapshot at least half a year old, so for the first six months those columns could only say that no snapshot exists yet. And many products sell for fewer than 26 weeks, so for them the columns would only ever say the forecast window exceeds the selling window. Once your account has six months of snapshots, ask your Toolio contact to turn the longer windows on. Everything in the popup — the table, the Average row, the banner and the chart — then extends to five windows.

The grid column blends the choice's clusters, weighted by actual sales units. The popup breaks that number apart.

Rows are clusters. Columns are grouped by forecast window, shortest first, and each group holds three cells:

Each column header is dated with the week the snapshot was taken, counted back from the last week measured. The subtitle above the table names that week, so the header dates always have an anchor.

The engine can differ between forecast windows on the same row. A row reading AI at four weeks and Statistical at thirteen simply means a different method was running a quarter ago, and the trend across that row spans both.

Most cells hold a plain percentage. These are the ones that do not, and what each is telling you.

Two different windows appear in this popup, and they are not related. The forecast window is the span of weeks a snapshot is scored over — it is what the columns are. The selling window is how long the offering is on sale — it is why a score can be impossible.

The difference between — and Exceeds the selling window is worth knowing. The first means not yet, and waiting will fix it. The second means never, because the offering's selling window is shorter than the forecast window.

Overall Forecast Accuracy at the top is the headline for the whole choice, taken at the shortest forecast window that has a score and weighted across clusters by actual sales units. It reads as a level, a percentage and a forecast window — High (94%) at 4 weeks — because the same percentage means very different things at four weeks and at a year.

These are the same bands as Forecast Confidence Score, so one scale reads across both popups. When nothing has been scored yet the banner reads — in grey rather than Low (0%).

The pinned Average row at the bottom of the table is the same calculation applied to each forecast window — weighted by actual sales units across the clusters that have a score. Clusters with no score are left out entirely rather than counted as zero.

Clusters that read no sales count in the average Accuracy at 0%, but they are left out of the average Bias. Their bias has no meaningful percentage, and including it would report an overforecast that never happened. If none of the scored clusters sold anything, the clusters are averaged equally instead of by sales.

The chart under the table plots both measures across the forecast windows shown in the table, shortest on the left. Read right to left to see whether the forecast got better as the forecast window shortened.

Use the selector above the chart to switch between All clusters (average) and a single cluster. The average matches the Average row exactly, and the row you are plotting is highlighted in the table. You can also click a cluster row to plot it, or click the Average row to go back.

A forecast window with no score anywhere leaves a gap rather than dropping to zero. If fewer than two forecast windows have been scored there is no trend to draw, and the chart says so instead of plotting a single point.

Because the 26- and 52-week forecast windows are hidden until they are turned on for your account. A 26-week score needs a snapshot at least half a year old, and many products never sell for 26 weeks, so on most accounts those columns would show the same waiting message on every row. Once you have six months of snapshots, ask your Toolio contact to enable them.

Because a 52-week score needs a snapshot that is a full year old, plus the year of sales that followed it. Long forecast windows fill a year after snapshotting started on your account, so this is the expected state until then.

Forecast Confidence Score estimates how much to trust a method before you plan, from a backtest. Forecast Accuracy reports how the forecast you actually used performed once sales came in.

Each cell is scored against the engine that served that snapshot. If the method changed during the quarter, the row shows both. The Forecast cell always names the engine behind the number beside it.

It means the cluster sold nothing in the measured weeks while the forecast expected units. Worth checking whether the cluster should have been forecast at all, or whether sales are landing somewhere else.

The column blends the choice's clusters, weighted by actual sales units. The popup shows each cluster separately, and its Average row usually matches the column.

They can differ when clusters have been scored at different forecast windows. The column takes each cluster at its own shortest scored forecast window, so it can mix a 4-week score from one cluster with an 8-week score from another. The banner and the Average row use one forecast window for every cluster.

No. Forecast Accuracy is read-only. It reports what happened and has no settings to adjust. To change how a choice is forecast, use Forecast Type in the Forecast Confidence Metrics popup.

Once a week. Toolio saves a new snapshot every week, and scores the day after your retail week closes, rescoring every forecast window whose sales have now arrived.

No. It covers all four weeks after the snapshot. Each forecast window is a horizon, not a lag. See A Forecast Window Is a Horizon, Not a Lag above.

No. Accuracy scores the System forecast the engine produced. Changes made to the forecast in the plan are not scored.

A forecast window is only scored when a snapshot exists for exactly that week. If a week's snapshot is missing, that forecast window is skipped for the runs that would have read it, and its last scored value stays on screen until it can be scored again.

It can. The selling window is read from the current launch and end dates, so shortening a window can put a longer forecast window permanently out of reach, and extending one can bring it back.

Forecast Confidence Score in New Assortment Plan

Forecast With AI in New Assortment Plan

Forecast in Assortment Plan