Sales & Volume Forecasting (Holt-Winters)

A time-series forecasting model that projects brand/category volume 12 months forward using Holt-Winters exponential smoothing — with an honest, back-tested accuracy check before it ever forecasts into the future.

---

What this does

1. Builds a monthly volume history with trend, seasonality, and noise
2. Splits the data into train/test and fits Holt-Winters on the training window only
3. Back-tests the fitted model against the held-out months and reports MAPE (Mean Absolute Percentage Error) — so the accuracy number is genuine, not measured on data the model already saw
4. Refits on the full history and forecasts the next 12 months forward
5. Outputs a chart (actual vs. forecast) and a CSV of the forward forecast

---

Repo structure

sales-volume-forecasting/
├── README.md
├── Sales_Volume_Forecasting.pyw
├── volume_forecast_next_12_months.csv
├── screenshots/
│   └── volume_forecast.png
└── LICENSE

Note: the forecast output was originally saved with an .xls extension, but its contents are plain CSV (Month,Forecast_volume). Rename it to .csv before uploading — Excel will otherwise flag it as an unrecognized/corrupt file when opened directly, even though it opens fine as text.

---

Methodology

1. Historical data

36 months (Jan-2023 – Dec-2025) of simulated monthly volume, built from three components:

trend       = linear growth from 100 to 130 (000s units)
seasonality = sinusoidal annual cycle + an extra Q4 holiday bump in December
noise       = random Gaussian noise (σ = 4)
volume      = trend + seasonality + noise

2. Train/test split & backtest

The last 6 months are held out as a test set. Holt-Winters (trend="add", seasonal="add", seasonal_periods=12) is fit on the remaining 30 months, then used to forecast those 6 held-out months. Comparing the forecast to the actual held-out values gives:

MAPE = mean(|Actual − Forecast| / Actual)

This run achieved a 3.7% backtested MAPE — the model's forecasts were off by less than 4% on average against data it never trained on, which is the number that should be quoted as the model's real-world accuracy (not an in-sample fit statistic).

3. Forward forecast

Once accuracy is validated, the same Holt-Winters specification is refit on the entire 36-month history (using all available signal) and used to forecast 12 months forward (Jan-2026 – Dec-2026).

Month	Forecast Volume (000s units)
Jan-2026	145.7
Feb-2026	140.9
Mar-2026	135.4
Apr-2026	133.6
May-2026	123.3
Jun-2026	122.0
Jul-2026	122.6
Aug-2026	128.3
Sep-2026	135.3
Oct-2026	141.8
Nov-2026	151.5
Dec-2026	162.3

The forecast shows a dip through mid-2026 followed by a strong Q4 ramp — consistent with the seasonal (holiday-driven) pattern learned from history.

---

Chart

Solid navy line = actual history · dashed gold line = 12-month forward forecast · title shows the backtested MAPE.

---

How to run

pip install numpy pandas matplotlib statsmodels
python Sales_Volume_Forecasting.pyw

This will:

• Print the backtested MAPE and the 12-month forecast table to the console
• Save the chart to volume_forecast.png
• Save the forward forecast to volume_forecast_next_12_months.csv

---

Tech notes

• Model: statsmodels.tsa.holtwinters.ExponentialSmoothing, additive trend + additive seasonality, 12-month seasonal period
• np.random.seed(42) makes the simulated history reproducible run-to-run
• Swap in real historical volume (indexed by month) in place of the simulated series to use this on actual data — the rest of the pipeline (backtest → MAPE → refit → forecast) needs no changes
• Because the seasonal component is additive, this fits data where seasonal swings stay roughly constant in absolute size; for volume series where seasonal swings scale with the trend level, switch to seasonal="mul"

---

License

This project is licensed under the MIT License — see LICENSE for details.