Understanding forecasts¶
This page explains what AccuPredix predicts and how to read the numbers — no
math background required. For engine internals, see the repository
model/MODEL_EXPLAINED.md.
What you get¶
For each product–customer combination (a series), AccuPredix produces a month-by-month demand forecast with an honest range of uncertainty.
Think of it like a weather forecast: not just "12 units," but "most likely 12, unlikely below 4 or above 20."
How the model chooses a method¶
There is no single best algorithm for every SKU. AccuPredix runs a competition of forecasting specialists on each series, scores them with walk-forward backtesting, and lets the winner forecast the future.
| Specialist | Best for |
|---|---|
| Seasonal-Naive | Same month last year |
| Moving Average | Stable, low seasonality |
| AutoETS | Clear trend and seasonality |
| Croston / TSB / ADIDA | Intermittent (many zero months) |
| LightGBM | Machine learning across all series |
| LightGBM + Indicators | ML plus external economic signals |
Demand personalities (smooth, erratic, intermittent, lumpy) route series to appropriate methods automatically.
Walk-forward backtesting¶
To score fairly, the model repeatedly pretends it is a few months in the past, forecasts forward, and compares to actuals. This walk-forward WAPE is the primary accuracy metric — it reflects real predictive skill, not memorisation.
Reading forecast bands¶
Each future month has three quantiles:
| Band | Meaning |
|---|---|
| P50 (q50) | Central / most likely estimate |
| P10 (q10) | Demand unlikely to fall below this |
| P90 (q90) | Demand unlikely to exceed this |
Bands are conformal-calibrated from backtest errors — they aim for correct coverage, not arbitrary confidence intervals.
Key metrics¶
| Metric | Simple view | Technical view |
|---|---|---|
| Accuracy | Accuracy score 0–100 | Portfolio WAPE (lower is better) |
| Series error | — | Best WAPE per series |
| Trend | Up / down / stable labels | % change vs trailing average |
| Data drift | Alert banner on dashboard | WAPE slope over last N runs |
WAPE (Weighted Absolute Percentage Error) measures forecast error relative to actual volume. An accuracy score of 85 roughly corresponds to strong WAPE; the exact mapping is shown in the app.
Overrides and scenarios¶
- Overrides replace the model output for specific months (logged and reversible).
- Scenarios apply temporary multipliers or templates for what-if planning without retraining.
Both affect displayed forecasts; exports can include or exclude overrides.
Cold start¶
Series with very little history use cold-start priors and are flagged as
estimates. Upload product_master with launch_date to improve new-SKU forecasts.
Data drift (Growth+)¶
When a series' backtest error trends upward across recent retrains, data drift alerts appear on the dashboard and in completion notifications — a signal that demand patterns are shifting faster than the model adapts. Consider an unscheduled retrain or reviewing recent events/promotions.
Safety stock¶
Growth+ workspaces can rank SKUs for replenishment using forecast quantiles and
optional margin data from product_master — see the Safety stock panel on
inventory-oriented reports.