analysis_plan.md

<!-- Prepared by Molla Samser (RSK World) | hello@rskworld.in | support@rskworld.in | +91 93305 39277 | Nutanhat, Mongolkote, Purba Burdwan, West Bengal, India 713147 | https://rskworld.in -->

# Analysis & Modeling Plan

Use this outline inside a Jupyter notebook to accelerate exploration.

## 1. Setup
- Import pandas, numpy, matplotlib/seaborn, scikit-learn, statsmodels, prophet (optional).
- Load `data/processed_sales.csv` (generated by `scripts/preprocess.py`).

## 2. Data Quality
- Inspect missing values, outliers, and negative quantities.
- Validate unique combos of `date`, `store_id`, `product_id`.
- Verify promo text consistency; derive binary promo flag.

## 3. Exploration
- Plot daily/weekly sales by category and region.
- Seasonal decomposition per top product (weekly/yearly).
- Price elasticity checks: scatter `price` vs `units_sold`.

## 4. Feature Engineering
- Calendar (day/week/month/quarter/year/day-of-year, holidays if available).
- Seasonal Fourier terms (sin/cos of day-of-year).
- Lag/rolling stats (7/14/28-day) and 4w/12w aggregates on `units_sold` and `revenue`.
- Price dynamics: price deltas, elasticity plots, revenue per unit.
- Inventory coverage: stock vs average daily demand.
- Categorical encodings for `product_category`, `region`, `customer_segment`.
- Promo flags, discount depth, and stock constraints.

## 5. Modeling
- Baselines: seasonal naïve, moving average.
- Tree/boosting: RandomForestRegressor, GradientBoosting/XGBoost/LightGBM/CatBoost with categorical support.
- Global time-series models: Temporal Fusion Transformer/DeepAR (if available) or Prophet/SARIMAX per product/segment.
- Cross-validation: expanding window; final hold-out on most recent period.

## 6. Evaluation
- Metrics: RMSE, MAE, MAPE; bias checks.
- Backtests per category/region to spot drift.

## 7. Packaging
- Persist trained models and scalers.
- Export forecasts with prediction intervals.
- Update `sales-forecasting.zip` with processed data, notebooks, and models.