forecasting

forecasting

When to use

• The annual plan or quarterly board pack needs a forecast model that survives a retro — not last quarter's number with a multiplier.
• Top-down (TAM × penetration × motion) and bottom-up (deal-level) calls have diverged and the reconciliation hasn't been written.
• A new finance-partner inherits a forecast and needs to rebuild the construction shape without inheriting the prior regime's optimism.

Do NOT use to qualify a single deal (route to deal-qualification-meddic), construct the RevOps commit list (route to forecast-accuracy (H10) — finance owns the shape, RevOps owns the call), or run capital-runway scenarios (route to runway-cognition (O3)).

Cognition cluster

• Mental model 9 — Hypothesis-driven thinking. Each forecast is a falsifiable claim about a window. If the call cannot be falsified inside the window, the call is a narrative, not a forecast. See mental-models.md § 9.
• Mental model 29 — Premortem. Before locking the call, write the post-window retro as if commit missed by 20 %. The premortem surfaces which construction inputs were riding on weak evidence; demote those before the call locks. See mental-models.md § 29.
• Mental model 16 — Leading vs lagging. Closed-won is lagging; pipeline coverage, segment conversion, and slot-completeness are leading. A forecast built only on lagging signals can confirm but not steer. See mental-models.md § 16.
• Context-spine — product + fiscal-period + customer-segment. Read the product slot for what is GA-shippable in the window; the fiscal-period slot for the cadence the model must reconcile against (monthly close vs quarterly board pack vs annual plan vs multi-year plan); the customer-segment slot for segment-historical close rates. See context-spine.

Procedure

Step 0: Inspect the construction shape

Read the fiscal-period slot. Decide between three shapes:

1. Top-down — anchor against TAM × penetration band × motion band. Healthy for annual plans and multi-year plans where bottom-up evidence is thin past one window.
2. Bottom-up — sum deal-level conviction (composes H10 forecast-accuracy via the forecast-construction-shape ADR). Healthy for quarterly windows where deal evidence is fresh.
3. Hybrid — both, with an explicit reconciliation. Healthy when top-down and bottom-up diverge by more than the historical confidence band.

State the choice. A forecast without a stated shape inherits the prior regime's shape silently.

Step 1: Construct the call against the shape

For top-down: write {tam, penetration_band, motion_band} — every input cites its source. Penetration bands are evidence ranges, not single points; motion bands reflect channel mix.

For bottom-up: consume H10's commit-list against the forecast-construction-shape interface. Sum commit-tagged × in-window close-rate per segment.

For hybrid: do both, then write the reconciliation. If top-down ≠ bottom-up by more than the confidence band, the divergence is the forecast — not either number.

Step 2: Calibrate the confidence band

Compute historical deviation from the last 4–8 windows of the same fiscal-period cadence. Attach as {plus_pct, minus_pct}. A band asymmetric on the downside is honest about prior misses; symmetric bands silently pretend prior accuracy.

Step 3: Premortem the construction

Write "if the forecast misses by 20 %, the reason is ___." For top-down: which penetration / motion input was the load-bearing assumption? For bottom-up: which anchor deals carry > 10 % of commit? Demote inputs that the premortem can name as single-point risks.

Step 4: Emit the typed interface

Produce forecast-band.json per the forecast-construction-shape ADR. H10 consumes the artifact for the commit-call. The fields: construction_shape, commit_value, best_case_value, pipeline_value, confidence_band, retro_signature, segment_scope, fiscal_period, construction_inputs. Drop the artifact in the location H10's ## Output references.

Step 5: Run the accuracy retro-loop

At window-end, compare predicted commit / best-case to actual closed-won. Compute per-segment and per-construction-input miss rate. Patterns that repeat for two windows become shape changes in Step 0 (e.g. switching from bottom-up to hybrid because deal evidence stopped predicting); one-off misses become input upgrades in Step 1.

Related Skills

WHEN to use this

• Constructing the finance-side forecast (annual plan, board pack, multi-year plan).
• Running the construction-shape retro and feeding it back into Step 0.

WHEN NOT to use this

• Single-deal qualification — route to deal-qualification-meddic.
• Commit / best-case / pipeline categorisation of deals — route to forecast-accuracy (H10); H10 consumes against this skill's forecast-band.json interface.
• Cash-runway shape and fundraise-trigger heuristics — route to runway-cognition (O3).
• Multi-statement scenario construction over base / upside / downside — route to scenario-modeling (O4).

Wing-4 handoff: this skill emits the forecast-band.json artifact that forecast-accuracy (H10, Wing-3) reads. Per docs/contracts/adr-forecast-construction-shape.md, docs/guidelines/wing4-handoff.md § Chain 4.

When the agent should load this

• "Build the annual forecast model."
• "Top-down and bottom-up disagree — reconcile them."
• "Why was last quarter's forecast off?"
• "Was machen wir bei der Forecast-Konstruktion anders?"

Output

1. forecast-band.json (Wing-3 / Wing-4 typed interface) — construction_shape, commit_value, best_case_value, pipeline_value, confidence_band, retro_signature, segment_scope, fiscal_period, construction_inputs. Per adr-forecast-construction-shape.md.
2. construction-notes.md — shape chosen + why; per-input evidence; reconciliation note (hybrid only).
3. premortem.md — "if we miss by 20 %, the reason is ___"; tagged demotions from Step 3.
4. retro-deltas.md (at window-end) — predicted vs actual per construction input; shape-change recommendation if the pattern repeats.

Gotcha

• A forecast without a stated construction_shape inherits last regime's shape silently. Always emit the field.
• Symmetric confidence bands lie about prior misses. If the last two windows missed on the downside, the band is asymmetric.
• Top-down models with single-point penetration assumptions are scenarios in disguise. Use bands.
• Hybrid models that don't write the reconciliation are top-down models with bottom-up garnish.

Do NOT

• Do NOT collapse hybrid forecasts into a single number without keeping the divergence visible.
• Do NOT skip Step 4 — the typed interface is what makes H10 reproducible.
• Do NOT change the construction shape on a single-window miss; shape changes require a two-window pattern.

Runnable example

End of FY: annual plan + Q1 commit both due.

• Step 0 — fiscal-period slot says annual + quarterly. Annual is top-down; Q1 is bottom-up.
• Step 1 — top-down: TAM $4.2B, penetration band 0.6–0.9 %, motion band SaaS-mid; expected $25–38M ARR. Bottom-up: H10 commit-list sums to $8.1M in Q1, segment close rate 78 %.
• Step 2 — last 4 quarters deviation: +6 % / –14 %. Confidence band attached.
• Step 3 — premortem: top-down anchored on penetration upper bound; demoted to 0.6–0.75 %. Bottom-up: two anchor deals tagged single-risk procurement; demoted.
• Step 4 — emit forecast-band.json: construction_shape=hybrid, commit $6.3M, best-case $8.1M, band +6/–14 %, retro_signature quarterly | [+6, –14], segment_scope mid-market, fiscal_period quarterly.
• Retro — at quarter-end, actual $6.1M; band held. Annual top-down revisit in two quarters.