defensible-forecastinglisted

Build defensible time-series forecasts from historical data (sales, GMV, demand, web traffic — any numeric series over time). ALWAYS interview the user first (goal, forecast horizon, output document format, charts yes/no, data structure) before doing anything else. Runs a bundled engine that backtests ~20+ methods across every major forecasting family (rolling-origin CV), selects the best 3 by MAPE with anti-overfitting safeguards (holdout confirmation, parsimony, seasonal data-gating), and outputs a forecast with prediction intervals and a plain-language risk rating. Use whenever the user wants to forecast, predict, or project future numbers from historical data — even when they only say "predict next month's sales", "project demand", "proyeksi penjualan", "peramalan", or "forecast". Trigger for any demand-planning, sales/target-setting, or time-series projection task, even if the word "forecast" is not used.
diarmilenia01-art/claude-skills · ★ 0 · Data & Documents · score 72

Install: claude install-skill diarmilenia01-art/claude-skills

# Defensible Forecasting Created by **Diar Azari** · https://www.linkedin.com/in/diarazari/ · 2026 · MIT License Produce forecasts you can defend: the chosen model wins on tested accuracy, and every result ships with its uncertainty and limits. ## Step 0 — Interview first (MANDATORY) Do NOT load, clean, or run anything until the user answers. Ask, in the user's language, then wait: 1. Goal — what to forecast, and for what decision? 2. Horizon — one period (1 day/week/month) or many (e.g. 30 days, 12 months)? 3. Output — Excel, Word, PDF, other? 4. Charts — yes or no? 5. Detail — concise or full? 6. Data structure — single series or many (e.g. per product/marketplace)? Or attach data to auto-detect. If told "just do it": defaults = horizon per data frequency, Excel, charts on, medium detail, auto-detect. Announce defaults, then proceed. ## Run the engine (do NOT rewrite forecasting code) Deps once: `pip install pandas numpy statsmodels scikit-learn openpyxl` ``` python scripts/forecast.py --input <data.csv|xlsx> --date-col <D> --value-col <V> \ --freq <D|W|M> --horizon <N> --output <forecast.xlsx> \ --lang <en|id> --unit "<prefix e.g. 'Rp '>" --label "<what the value is>" \ [--folds 8] [--holdout 6] [--pi 80] ``` The script does: profiling → quality check → aggregation → seasonal data-gating → rolling-origin backtest → stability filter (drops numerically unstable models) → safeguarded selection (holdout + parsimony) → top-3 + ensemble → Excel report whose **fir