Merchant Revenue Forecasting (v1)

Goal: Forecast monthly merchant revenue across 50 merchants (5 years of data) and compare multiple forecasting methods.
Why: Advanced forecasting reduces error in budgets and informs resource planning.
Result: Advanced models (Holt-Winters, SARIMA, Linear Regression) cut forecast error from ~12–13% (Seasonal Naive baseline) to ~8–10% MAPE.

---

Dataset

• Scope: 50 merchants × 60 months (Jan 2020 – Dec 2024).
• Columns: date, merchant_id, region, vertical, revenue, orders, avg_order_value, marketing_spend, promo_month, seasonal_index, macro_index, mktg_rev_ratio, cac_per_order.
• Features:
  • Seasonal Index: captures within-year demand swings (e.g., Dec > 1 = holiday boost; Jan < 1 = lull).
  • Macro Index: economy-wide effects (shock + recovery in 2020).
  • Marketing spend with diminishing returns, plus promo lifts.

Synthetic data was generated but designed to mimic realistic FP&A challenges.

---

Methods

• Baselines: Naive, Seasonal Naive, Simple Moving Average (SMA), Weighted Moving Average (WMA).
• Classical: Simple Exponential Smoothing (SES), Holt’s Linear Trend, Holt-Winters Seasonal.
• Time Series: SARIMA.
• Drivers: Linear Regression with marketing spend, promo flags, seasonality, and lagged revenue.

---

Evaluation

• Train/Test: Last 12 months per merchant used for testing.
• Metrics: MAPE, sMAPE, MAE, RMSE, MASE.
• Portfolio View: Results aggregated across all merchants into leaderboards and distribution charts.

---

Results

Average Forecast Error (MAPE)

• Baseline (Seasonal Naive): ~12–13% MAPE
• Best advanced models: Holt-Winters / SARIMA / Linear Regression ~8–10% MAPE
• ~30% relative improvement in forecast accuracy

---

Distribution of Errors (Boxplot)

• Each box = error spread across merchants.
• Orange line: median error.
• Green triangle: mean error.
• Whiskers/dots: variation and outliers.
• Advanced models show both lower medians and tighter distributions → more consistent across merchants.

---

Example Merchant Forecast

• Actual vs forecast for one merchant.
• Shows improvement of Holt-Winters/SARIMA over the naive baseline, especially in capturing seasonality.

---

Next Steps

Next (V2): Incorporate promotions, marketing spend and macro index as external regressors in SARIMA and advanced ML models (XGBoost, LightGBM).

---

Business Implications

• Tighter planning: Cutting error from ~12–13% to ~8–10% improves budgets and inventory/cash planning.