back
loading skill details...
Interrogate a demand forecast before the business commits supply and inventory to it. Use when asked to review a demand plan, challenge a forecast, check for...
---
name: demand-forecast-review
description: "Interrogate a demand forecast before the business commits supply and inventory to it. Use when asked to review a demand plan, challenge a forecast, check forecast accuracy, decompose baseline vs uplift, or find hockey sticks in the numbers. Produces a forecast credibility review with baseline/uplift decomposition, MAPE and bias history, hockey-stick flags, an assumption register, and consensus-vs-statistical divergence analysis."
homepage: https://mohitagw15856.github.io/pm-claude-skills/skill/demand-forecast-review.html
metadata:
{
"openclaw": { "emoji": "📦" }
}
---
# Demand Forecast Review Skill
Every unit of forecast becomes purchase orders, capacity commitments, and inventory. This skill interrogates a forecast the way a supply planner must: separate the defensible baseline from hopeful uplift, confront the forecast with its own accuracy history, hunt for hockey sticks, and register every assumption so that when the number misses, you know which belief broke.
## What This Skill Produces
- A baseline vs. uplift decomposition (statistical base + named uplift layers)
- Forecast accuracy history: MAPE and bias, with what they imply for buffering
- Hockey-stick and pattern-anomaly flags
- An assumption register with owner, evidence strength, and expiry
- Consensus vs. statistical divergence flags with a burden-of-proof call
- An overall credibility verdict: plan to it / plan to it with buffers / send back
## Required Inputs
Ask for these if not provided:
- **The forecast** — by product/family and period, over the horizon under review
- **History** — actuals for the trailing 12+ months; prior forecasts vs. actuals if available (for MAPE/bias)
- **Uplift drivers** — promotions, launches, new customers, pipeline deals baked into the number
- **Who built it** — statistical, sales-driven, consensus; and what changed since last cycle
- **Decision at stake** — what the forecast will commit (buy, build, capacity) and its lead time
With no accuracy history, review structure and assumptions and state plainly: `[accuracy unknown — treat forecast as unvalidated]`. Never present conclusions as if history existed.
## Interrogation Framework
**1. Decompose baseline vs. uplift.** Baseline = what history alone supports (trend + seasonality). Everything above it is an uplift layer that must be named: which promotion, which customer, which launch. Compute uplift share of total — above ~30% uplift, the forecast is a sales plan wearing a forecast's clothes, and each layer needs its own evidence.
**2. Confront accuracy history.**
| Metric | Read it as | Action threshold |
|---|---|---|
| MAPE (lag matched to decision lead time) | Noise level | >30% at family level: forecast can't carry item-level commitments |
| Bias (signed error, running) | Systematic lean | Same sign 3+ consecutive periods: correct the input, don't buffer around it |
Persistent over-forecast bias means excess inventory is being manufactured upstream; persistent under-forecast means service failures are planned in. Name which one this forecast has.
**3. Hunt hockey sticks.** Flag: quarter-end/year-end spikes with no order-book support; growth rates that jump beyond trailing actuals precisely when the plan needs them to; a ramp that has slipped right by one quarter in each successive cycle (the sliding hockey stick — the strongest sell-back signal there is).
**4. Register assumptions.** Every uplift and step-change gets a row: assumption, owner, evidence (order book / customer commitment / pipeline / hope), the period when reality will confirm or kill it, and the volume at stake if it fails.
**5. Flag consensus vs. statistical divergence.** Where consensus overrides the statistical line by >10%, the override carries the burden of proof. Check the track record: have past overrides beaten the stat model? If overrides historically added error, recommend planning supply to the statistical line and treating the delta as upside to option, not to stock.
## Output Format
### Forecast Review: [scope / cycle]
**1. Verdict** — plan to it / plan with stated buffers / send back for rework, and the one-paragraph why.
**2. Decomposition** — table: Period | Baseline | Uplift layer(s) | Total | Uplift %.
**3. Accuracy history** — MAPE and bias at the decision lag, trend, and the buffering implication.
**4. Flags** — hockey sticks, sliding ramps, anomalies vs. history, each with the volume at stake.
**5. Assumption register** — Assumption | Owner | Evidence strength (committed / probable / speculative) | Confirms by | Units at stake.
**6. Divergence analysis** — consensus vs. statistical by family; where overrides exceed 10%, the recommendation on which line supply should plan to.
**7. Questions for the demand owner** — the 3–5 questions that must be answered before commitment.
## Quality Checks
- [ ] Baseline and uplift separated, with uplift % computed and every layer named
- [ ] Accuracy metrics computed at the lag matching the commitment lead time — or absence declared
- [ ] Bias reported as signed and directional, not folded into MAPE
- [ ] Every hockey-stick flag cites the pattern evidence (order book, prior-cycle slippage)
- [ ] Every speculative assumption has an owner and a confirm-by date
- [ ] Verdict states what supply should actually plan to, not just critique
## Anti-Patterns
- [ ] Do not present a forecast without its historical accuracy — a number with no track record is a guess with a spreadsheet
- [ ] Do not treat persistent bias as noise to buffer — a lean that repeats is an input error to fix at source
- [ ] Do not let uplift hide inside the baseline — unnamed uplift is unaccountable uplift
- [ ] Do not accept "the ramp moved right but the year is intact" without flagging it — sliding ramps rarely land
- [ ] Do not judge accuracy at aggregate level for item-level buys — mix error is where the money is lost
- [ ] Do not soften the verdict to keep the S&OP meeting comfortable — supply commits real cash to this number
don't have the plugin yet? install it then click "run inline in claude" again.
extracted and formalized 6-component structure from interrogation framework, added decision points for missing history and bias persistence, clarified external connection inputs, expanded output contract with explicit table schemas and quality checklist, added outcome signal tied to s&op readiness.
interrogate a demand forecast the way a supply planner must before the business commits purchase orders, capacity, and inventory to it. use this skill when asked to review a demand plan, challenge a forecast number, check forecast accuracy against history, decompose what's defensible baseline versus what's hopeful uplift, hunt for hockey sticks and anomalies, or validate assumptions buried in the plan. output: baseline/uplift decomposition, MAPE and bias history with buffering implications, hockey-stick flags, a named assumption register, and consensus-vs-statistical divergence analysis. verdict states plainly whether to plan to the number, plan with stated buffers, or send it back.
forecast data
accuracy history
uplift drivers
forecast provenance
decision context
external connections (if available)
if no accuracy history exists, proceed to decomposition and assumption validation, but flag plainly: [accuracy unknown , treat forecast as unvalidated]. never present conclusions as if history existed.
step 1: decompose baseline vs. uplift
step 2: compute accuracy metrics at commitment lead time
[accuracy unknown]. if MAPE > 30% at family level, flag: forecast insufficient for item-level commitments. if bias is same sign for 3+ consecutive periods, flag: systematic input error, not noise; recommend correcting the input, not buffering around it.step 3: hunt hockey sticks and anomalies
step 4: register assumptions
step 5: flag consensus vs. statistical divergence
step 6: verdict and recommendation
step 7: generate questions for demand owner
if no accuracy history is available: proceed to decomposition and assumption validation, but declare accuracy unknown and do not cite historical MAPE/bias to justify buffering. treat forecast as unvalidated input.
if MAPE > 30% at family level: do not recommend item-level commitments against this forecast. recommend family-level or aggregate positioning, or recommend statistical model refit and resubmission.
if bias is persistent and same-signed (3+ consecutive periods): do not recommend buffering around the bias (e.g., over-forecasting with inventory buffer). instead, recommend fixing the input (e.g., adjust the forecast method, correct a data quality issue, or re-baseline). systematic bias is an input error, not noise.
if uplift exceeds 30% of total forecast: each uplift layer requires independent evidence (order book, customer commitment, or signed pipeline). if any layer is speculative or unowned, flag for resolution before commitment.
if hockey-stick detected without order-book support: flag as high-risk. if order book exists but forecast growth exceeds order growth, question the delta explicitly. if ramp has slipped right in successive cycles, recommend planning to a conservative baseline and treating the ramp as optionality.
if consensus override exceeds 10% vs. statistical line: check historical track record of that override rule. if past overrides increased error, recommend planning to statistical line. if past overrides beat the model, document the rule and proceed. if no track record, treat override as speculative and require explicit owner and confirm-by date.
if assumption confirm-by date has passed without validation: forecast is stale and should not be used for new commitments. demand owner must re-validate or retract the assumption.
if forecast will commit supply with lead time > 12 weeks and accuracy history is sparse: flag as high-risk. recommend splitting commitment into committed (validated) portion and option (validated at closer date).
output is a structured forecast review document in markdown or table format. must contain these sections in order:
1. verdict , one-sentence recommendation (plan to it / plan with buffers / send back), followed by one-paragraph rationale.
2. decomposition table , columns: period, baseline (units), uplift layer 1 (units), uplift layer 2 (units), ..., total forecast (units), uplift % of total. show at least 8 periods. baseline must be separated from each named uplift layer.
3. accuracy history , if history exists: MAPE (family level), bias (signed, by period for trailing 8+ cycles), trend in both metrics, confidence interval or range. include lag (e.g., "forecast lag: 8 weeks"). if history absent: declare [accuracy unknown , treat as unvalidated]. include buffering implication: e.g., "MAPE 22% implies ±22% buffer at 68% confidence" or "persistent over-forecast bias of 8% implies historical inventory excess; do not add buffer."
4. flags , list of hockey sticks, anomalies, ramp slippage, structural breaks. each flag must cite pattern evidence (e.g., "Q3 spike 45% above Q2, no order-book support"; "ramp slipped right by 1 quarter in each of last 3 cycles"). include volume at stake for each flag.
5. assumption register , table: assumption | owner | evidence strength (committed/probable/speculative) | confirm-by date | units at stake. do not allow unnamed assumptions; do not allow blank confirm-by dates for speculative assumptions.
6. divergence analysis , if consensus overrides statistical baseline: table showing statistical baseline, consensus forecast, delta (%), and decision rule / track record. recommendation on which line supply should plan to.
7. questions for demand owner , 3-5 numbered questions, each with a clear decision implication. e.g., "if new customer X does not sign by [date], what is fallback forecast? (decision: commit to [baseline] or [buffer] pending signature.)"
8. quality checklist , explicit pass/fail on: baseline and uplift separated; MAPE/bias computed at matching lag or absence declared; bias reported as signed; every hockey-stick cites evidence; every speculative assumption has owner and confirm-by date; verdict states what supply should plan to.
forecast review is complete and ready for s&op/demand-supply review when:
demand owner must sign off on (a) assumption ownership and confirm-by dates, or (b) accept the rework request, before forecast enters supply commitment.