This playbook teaches a deliberate division of labor: statistical models forecast, LLMs critique.
Tier: Core · Hands-on time: ~30 minutes
LLMs are poor time-series forecasters — they pattern-match plausible curves and cannot be held to error bars — but they are excellent at interrogating a forecast's assumptions, spotting missing drivers, and stress-testing the story you are about to take into a capacity or budget review. Use the AI as the hostile reviewer your forecast never gets.
Do NOT use this playbook to have an LLM produce the forecast numbers. If you have no statistical forecast yet, build one first — even a simple seasonal-naive baseline — then come back.
State the decision and the asymmetry. Write one sentence: "This forecast decides X; overestimating costs us Y, underestimating costs us Z." Everything the AI critiques is judged against this asymmetry, not abstract accuracy.
Feed the forecast, not the request for one.
You are a skeptical capacity-review board member. Attached: 18 months of
weekly peak utilization, a 6-month forecast produced by [method], and a
list of known upcoming events.
Do NOT produce your own forecast or adjust the numbers.
Critique this forecast:
1. Assumptions: what must be true for this projection to hold? List each.
2. History checks: seasonality, trend breaks, or one-off spikes in the
history that the method may have mis-learned. Point to specific dates.
3. Missing drivers: which listed events are not reflectable in the history?
4. Failure modes: 3 concrete scenarios where reality beats the forecast
high, 3 where it comes in low, each with an early-warning metric.
5. Questions a reviewer should ask the author before approving.
Answer the critique in writing. For each assumption and question, write the actual answer. Where you cannot answer ("we don't know if the client migration lands in Q3"), that becomes an explicit scenario branch in the forecast, not a silent hope.
Turn failure modes into tripwires. For each credible "reality beats forecast" scenario, define the early-warning metric and threshold (e.g., "if weekly peak grows >6% for 3 consecutive weeks, re-forecast"), and wire it as a real alert. This converts a static forecast into a monitored one.
Stress the story, not just the curve. Ask a follow-up: "Given the decision costs stated, argue the strongest case that we are about to over-provision, then the strongest case for under-provisioning." Read both; the side that makes you uncomfortable is where to spend one more hour of analysis.
Package for review. Have the AI condense into a one-page brief: forecast + method, assumptions register with answers, scenario branches, tripwire alerts. This is what goes to the capacity/budget meeting.
If the AI's critique changes the numbers, change them by re-running the statistical model with corrected inputs — never by letting the LLM "adjust" the projection.
The most expensive forecasting failure mode is a confident LLM curve with no backtest behind it — this playbook exists so that never reaches a purchase order.