菜单
Use when you want the model to learn from how its past forecasts actually turned out - scoring forecasts against the company's real actuals, backtesting assumptions on history, and proposing ratified improvements. Runs at/after monthly close.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when the user wants their model in Excel with real working formulas - "export this to Excel", "a workbook I can hand my board", "something that recalculates when I change an assumption". Produces a live-formula workbook generated from the model structure and verified against the engine before delivery.
Use when producing a board deck, investor update, or CFO briefing from an openfpa forecast - turning model output into a board-ready narrative and exportable artifacts.
Use when building a new openfpa forecast model from a company's financials - a trial balance, a P&L export, or a pasted income statement - and you need a runnable config to exist before any forecasting or analysis.
Use when modeling a multi-segment company that discloses segment net sales and segment Adjusted EBITDA (ASU 2023-07) but not segment COGS or operating income - rolls segment P&Ls into a consolidated forecast and reconciles total segment Adjusted EBITDA to the disclosed total. Generated for Fox Factory (PVG/AAG/SSG).
Use when a human reviewing a forecast catches something off ("December always spikes", "you're double-counting deferred revenue", "that Q3 number was a one-time contract") - captures it as a durable, typed correction in the company's memory so future forecasts are grounded by it.
Use when answering "when do we run out of cash", building a 13-week cash forecast, sizing a credit line, or analyzing near-term liquidity and payment timing in openfpa.
| name | fpa-backtest-learn |
| description | Use when you want the model to learn from how its past forecasts actually turned out - scoring forecasts against the company's real actuals, backtesting assumptions on history, and proposing ratified improvements. Runs at/after monthly close. |
The model should get measurably better at this business over time. This skill
scores past forecasts against the company's actuals, surfaces what keeps missing,
and proposes improvements a human ratifies. The objective metric is reconciliation
error against the user's own books (pyfpa.score_forecast) - the FP&A analog of a
validation loss.
Core principle: self-experimenting, but never self-promoting. The AI may run
and discard bounded challengers autonomously; a human approves replacement of
the champion. Everything learned lives as plain files in .fpa/.
.fpa/)forecasts/<period>.snapshot.yaml - each forecast's assumptions + predictions, and (after close) its score.scorecard.md - the running track record (rendered, never hand-edited).experiments/<slug>.experiment.yaml - each tested model change, its evidence,
changed files, checks, before/after metrics, and decision.learnings.md - every accepted change: what, the evidence, the backtest delta, the date.snapshot_forecast(cfg, forecast_df, label=<period>, created=<today>) →
save_snapshot(..., ".fpa/forecasts/<period>.snapshot.yaml").score_forecast(snap.predicted, actuals), write the
score back into the snapshot, and re-render scorecard.md with render_scorecard.type: parametric correction's target line keeps missing, flag it as possibly stale (applied → superseded) for the human - never auto-revert.pyfpa.save_experiment.holdout_backtest on the
company's history. Surface it only if it lowers holdout fitness (not in-sample),
ranked by the delta. Clamp the proposed move with magnitude_cap (±25%/cycle).persistent_miss is true for the line (same-signed across K≥2 closes)
and it survived the one-time screen. Hand it to fpa-learn-business to generate the
skill on approval - propose, don't auto-write.ExperimentDecision, set status: accepted, save with explicit
overwrite=True, update the company model, and append to learnings.md.
Rejected and reverted experiments remain in memory.fpa-research-loop when the miss warrants multiple autonomous
challenger epochs instead of one manually proposed change.If the company already has ~12+ months of actuals, you don't have to wait: run
holdout_backtest immediately (fit on the earlier months, hold out the recent ones)
to report current accuracy and the first round of parametric proposals.
holdout_backtest enforces this).magnitude_cap).Scored + learned → fpa-board-briefing (report the forecast and how it's tracking).