AI Prompts for Demand Forecasting

A demand forecast is only as useful as the reasoning behind it. Most forecasting problems in operations are not about the arithmetic — spreadsheets and planning systems do that — but about choosing a method that matches the data, separating baseline demand from one-off events, and writing down the assumptions so the forecast can be challenged next month.

AI is useful at three points: picking an approach from the shape of your history, producing a reasoned baseline with ranges rather than a single number, and structuring the judgmental overlay that sales and marketing add on top. It is not a substitute for your planning system's statistical engine, and it should never invent history you have not given it.

Before you use these

Have these ready to replace the highlighted [variables]:

• 12–36 months of sales or shipment history by SKU, family or channel
• Known events in that history (promotions, stockouts, price changes, launches)
• Planned events for the forecast horizon
• The granularity and horizon the forecast needs to serve (weekly for 13 weeks, monthly for 18 months)

The prompts

• 1. Choose a forecasting approach from the data
• 2. Build a baseline forecast with an assumptions log
• 3. Structure the judgmental overlay for events
• 4. Forecast a new product from analogs

1. Choose a forecasting approach from the data

You are a demand planner advising on forecasting method selection. Do not forecast yet.

Product / family: [product or family]
History: [paste series with dates, or describe: months of data, typical volume, min/max]
Forecast needed: [granularity] for [horizon]
Known drivers: [seasonality, price, promotions, weather, channel mix]

Work through this in order:
1. Characterize the series: trend, seasonality (period and strength), intermittency (share of zero periods), volatility (coefficient of variation), structural breaks.
2. Identify periods that should be cleansed or flagged before fitting (stockouts, one-off promotions, launch ramp) and say what evidence you would need to confirm each.
3. Recommend one primary method (e.g. seasonal exponential smoothing, moving average, croston/intermittent, regression on drivers, or judgmental) and one fallback. Justify each against the characteristics in step 1.
4. State what history length and aggregation level the method needs, and whether forecasting at family level then disaggregating would be more reliable.
5. List the information you do not have that would change the recommendation.

Do not invent data points. If the series is too short or too sparse for any statistical method, say so and recommend a judgmental or analog approach instead.

2. Build a baseline forecast with an assumptions log

Act as a demand planner building a baseline forecast for review.

Series (cleansed of known anomalies): [paste periods and values]
Method to apply: [method] — apply it transparently and show the calculation logic.
Horizon: [n periods] at [granularity]
Assumptions to apply: [e.g. underlying growth 3% YoY, seasonality as prior two years, no price change]

Produce:
A. Baseline forecast table: period, base, low, high. Derive low/high from historical error or stated volatility — state which.
B. Assumptions log: each assumption, its source (data, stakeholder, judgment), and the direction the forecast moves if it is wrong.
C. Seasonality profile used, expressed as indices, and the periods where the index is least reliable.
D. The three assumptions with the largest effect on the total, with the total forecast under each reversed.
E. What you were not given that would materially change the numbers (e.g. open orders, planned promotions, new distribution).

Constraints: keep all arithmetic visible enough to be audited. Do not extrapolate beyond the horizon. Do not smooth over a structural break — flag it and forecast from the post-break level if the break is confirmed.

3. Structure the judgmental overlay for events

You are facilitating the consensus step of a demand review. The statistical baseline is fixed; your job is to structure the judgmental overlay.

Baseline: [paste baseline by period]
Planned events: [for each: event, period(s), owner, expected effect if known]
Historical analogs: [past events with their measured lift or loss, if any]

For each event:
1. Estimate the adjustment (units or %) using the closest analog. If no analog exists, say so and give a range with the reasoning.
2. Classify the evidence: measured analog / stakeholder estimate / assumption.
3. Note the timing effect: pull-forward, post-event dip, cannibalization of other items.
4. Identify who owns the assumption and what would confirm it before the period starts.

Then produce:
- Overlay table: period, baseline, adjustment, consensus, evidence class, owner.
- Total baseline vs total consensus, and the share of the change coming from unmeasured assumptions.
- The overrides you would challenge in the review, with the question to ask.

Do not apply an adjustment you cannot attribute to an event. Do not compound overlapping events without stating the interaction assumption.

4. Forecast a new product from analogs

Act as a demand planner forecasting a new product with no sales history.

New product: [description, price point, target segment, channels and number of doors/accounts at launch, marketing support, expected cannibalization of existing items]
Analogs: [for each: product, why comparable, launch conditions, sales by period for the first 6–12 periods, peak period, share of year-1 volume in first quarter]
Constraints: [supply available at launch, MOQ, shelf life]

1. Rate each analog for comparability on segment, price, distribution breadth, marketing support and seasonality of launch date. Weight them accordingly and explain any analog you exclude.
2. Build a normalized launch curve (share of year-1 volume per period) from the weighted analogs.
3. Estimate year-1 volume using a driver approach (doors × velocity per door, or category share × category volume) and reconcile it against the analogs' absolute volumes. Show both and the gap.
4. Apply adjustment factors for differences from the analogs (distribution breadth, price, support) — state each factor, its direction and size, and whether it is evidence-based or judgment.
5. Produce low/base/high by period and the assumptions that separate the cases.
6. Set review points: after which period actuals should replace the analog curve, and the signal that would trigger an early revision (e.g. sell-through below the low case in period 2).
7. Supply implication: what the high case requires versus what is available.

Be explicit that the forecast is analog-driven; do not present precision the method cannot support.

Worked example

Related prompts

• Forecast Accuracy Review
• Sales & Operations Planning (S&OP)
• Safety Stock Calculation
• Capacity Planning
• ABC and ABC-XYZ Inventory Analysis
• Sales forecasting — Sales

Logical next step

After this, most operations teams move on to Forecast Accuracy Review.

All Operations & Supply Chain prompts · Search the full library