AI Prompts for Forecast Accuracy Review

Measuring forecast accuracy is easy; understanding it is not. A single MAPE hides whether the error is random noise, a systematic bias in one direction, a handful of large items dragging the average, or a process problem such as late overrides. The review that matters asks where the error comes from and whether each step in the process is adding value or subtracting it.

These prompts follow that sequence: compute the metrics properly at the right level, decompose error into bias and variance by segment, and then assess forecast value added so you know whether the statistical baseline, the planner overrides or the sales overlay are improving the number.

Before you use these

Have these ready to replace the highlighted [variables]:

• Forecast and actuals by item and period, ideally with the forecast as it stood at a fixed lag (e.g. one month out)
• Item attributes for segmentation (ABC class, family, channel, lifecycle)
• If available, the forecast at each process step (statistical, planner, consensus)

The prompts

• 1. Calculate and interpret accuracy metrics
• 2. Diagnose systematic bias by segment
• 3. Assess forecast value added and recommend process changes

1. Calculate and interpret accuracy metrics

Act as a demand planning analyst reviewing forecast accuracy.

Data: [paste item, period, forecast, actual — note the forecast lag used]
Segments available: [ABC class / family / channel / lifecycle]

Steps:
1. Compute per item and per segment: absolute error, MAPE (excluding zero-actual periods, and report how many were excluded), WAPE (volume-weighted), bias (signed error / actual), and a tracking signal (cumulative error / MAD).
2. State which metric should be the headline for this dataset and why (e.g. WAPE for a volume-driven business with many low-volume items).
3. Identify the ten items contributing most to total absolute error and their share of total volume.
4. Flag any segment where bias exceeds ±[threshold, e.g. 10%] — this indicates systematic over- or under-forecasting rather than noise.
5. Give the interpretation in plain language for a planning manager: what is good, what is a problem, and what is simply the nature of the demand.

Do not calculate accuracy at a level the data does not support. If the forecast lag is not stated, ask for it — accuracy at lag 0 is meaningless.

2. Diagnose systematic bias by segment

You are diagnosing why forecast bias is persistent in specific segments.

Bias findings: [segment, bias %, number of periods, direction]
Process context: [who forecasts, override policy, any incentives, recent changes]
Events in the period: [promotions, stockouts, new customers, price changes]

For each biased segment, work through candidate causes:
- Statistical: model type unsuited to the pattern, trend lag, seasonality mis-specified, outliers not cleansed.
- Process: overrides consistently in one direction, late-stage adjustments, forecast frozen too early.
- Behavioral: incentive to over-forecast (to secure supply) or under-forecast (to beat target).
- Structural: lost or gained customers not reflected, cannibalization, channel shifts.

Output:
1. For each segment, the two most likely causes, with evidence for and against from the data and context given.
2. A test for each: what to compare or measure to confirm it within one cycle.
3. Whether the bias is correctable by a method change, a process change, or a conversation — and with whom.

Do not attribute bias to individuals; describe the mechanism. If the data cannot distinguish between causes, say which additional field would separate them.

3. Assess forecast value added and recommend process changes

Act as a forecasting process consultant applying Forecast Value Added (FVA) analysis.

Data: [per item/period: naive forecast, statistical forecast, planner-adjusted forecast, consensus forecast, actual]
Process description: [steps, participants, time each step takes]

1. Compute accuracy (WAPE and bias) for each stage and for the naive forecast.
2. Calculate FVA for each step as the accuracy improvement versus the prior step. Identify steps with negative FVA — where effort makes the forecast worse.
3. For each negative or negligible step, explain the likely reason (over-adjustment, anchoring, late information not actually informative) and whether the step should be removed, restricted to exceptions, or refocused.
4. Recommend a revised process: which items get statistical-only forecasts, which warrant judgmental review, and what threshold should trigger a manual override.
5. Estimate the planner time saved and the accuracy effect, with the assumptions behind both.

Keep the recommendations proportionate to the evidence. One month of data does not justify removing a step; say what sample size would.

Related prompts

• Demand Forecasting
• Sales & Operations Planning (S&OP)
• ABC and ABC-XYZ Inventory Analysis
• Operational KPI Reviews
• Sales forecasting — Sales

Logical next step

After this, most operations teams move on to Demand Forecasting.

All Operations & Supply Chain prompts · Search the full library