This is a self-directed finance analytics / forecasting portfolio project that benchmarks nine monthly revenue forecasting methods across a synthetic merchant portfolio.
- 50 synthetic merchants
- 60 months of history (January 2020 – December 2024)
- 9 forecasting methods evaluated on a common holdout
- Holt produced the lowest mean MAPE at approximately 8.6%
- Seasonal Naive baseline mean MAPE was approximately 12.5%
- That is approximately a 31% relative reduction in mean MAPE versus the Seasonal Naive baseline
The dataset is fully synthetic. It is designed to resemble realistic FP&A forecasting challenges and does not contain real merchant, employer, or customer data.
How can Finance improve monthly merchant revenue forecasting across a heterogeneous merchant portfolio, and which forecasting method provides the best balance of accuracy, stability, and interpretability?
In an FP&A setting, the practical decision is which method to prefer for merchant-level planning support: a simple seasonal baseline, a classical smoothing approach, a seasonal time-series model, or a driver-based regression model.
- Type: Synthetic panel dataset (no real-world merchant identities)
- Coverage: 50 merchants × 60 months = 3,000 observations
- Period: January 2020 through December 2024
- Grain: One row per merchant-month
- Primary target:
revenue - Key fields:
date,merchant_id,region,vertical,revenue,orders,avg_order_value,marketing_spend,promo_month,seasonal_index,macro_index,mktg_rev_ratio,cac_per_order - Data dictionary:
data/raw/merchant_monthly_data_dictionary.json - Raw file:
data/raw/merchant_monthly_revenue.xlsx - Generator:
data/generator/data_generator.pywith fixed seed42for reproducibility
The portfolio pipeline (source/all_merchants.py) benchmarks nine methods:
| Group | Methods |
|---|---|
| Baselines | Naive, Seasonal Naive, SMA(3), WMA(1,2,3) |
| Classical smoothing | SES, Holt, Holt-Winters |
| Time series | SARIMA |
| Driver-based | Linear Regression (marketing spend, promo flag, macro index, lagged revenue, month effects) |
A single-merchant example script produces an illustrative actual-vs-forecast chart for merchant M001. That example is intentionally lighter than the full nine-model portfolio benchmark.
- Horizon: 12-month-ahead forecast
- Train / test split: For each merchant, the final 12 months are held out for testing; earlier months are used for training
- Primary comparison metric: MAPE (mean absolute percentage error), useful for comparing relative error across merchants of different scale
- Supporting metrics: RMSE, MAE, sMAPE, and MASE
Multiple metrics are reported because no single error measure is universally best. MAPE is convenient for portfolio comparison, while RMSE/MAE emphasize absolute error and MASE provides a scale-free check relative to a seasonal naive benchmark.
Verified mean MAPE ranking across all 50 merchants:
| Rank | Model | Mean MAPE |
|---|---|---|
| 1 | Holt | ~8.64% |
| 2 | SES | ~9.08% |
| 3 | SMA(3) | ~9.34% |
| 4 | Holt-Winters | ~9.54% |
| 5 | WMA(1,2,3) | ~9.60% |
| 6 | Linear Regression | ~9.66% |
| 7 | Naive | ~11.68% |
| 8 | SARIMA | ~11.81% |
| 9 | Seasonal Naive | ~12.47% |
Headline result: Holt achieved the lowest mean MAPE (~8.6%) versus Seasonal Naive (~12.5%), an approximate 31% relative reduction in mean MAPE.
Governance takeaway: Greater model complexity did not guarantee better average performance. SARIMA, for example, was not among the strongest average performers in this benchmark. Method selection should be evidence-based and refreshed as new observations arrive.
Best-model counts also vary by merchant (Holt was most frequently best by MAPE, followed by Linear Regression and Holt-Winters), reinforcing that portfolio averages and merchant-level results should be reviewed together.
Lower and more stable merchant-level forecast error can support:
- budgeting and revenue planning
- resource allocation discussions
- scenario planning around uncertain merchants
- identifying where forecast uncertainty is persistently higher
This project does not estimate dollar savings, claim production deployment, or assert real employer outcomes. The value demonstrated here is methodological: a transparent, reproducible framework for comparing forecasting approaches before using them in planning workflows.
Portfolio-level comparison of mean MAPE by method. Holt is lowest (~8.6%); Seasonal Naive is the weakest baseline (~12.5%).
Spread of merchant-level MAPE by method. Useful for assessing consistency, not only average error.
Actual versus forecast paths for merchant M001 over the 12-month test window (illustrative single-merchant view).
data/
generator/data_generator.py # synthetic data generator (seed=42)
raw/merchant_monthly_revenue.xlsx # synthetic panel dataset
raw/merchant_monthly_data_dictionary.json
transformed/ # generated Excel outputs (not committed)
functions/
forecast_methods.py # forecast helpers + horizon H=12
metrics.py # MAPE, RMSE, MAE, sMAPE, MASE
source/
all_merchants.py # full 9-model portfolio benchmark
single_merchant_forecast.py # illustrative M001 forecasts
single_merchant_forecast_metrics.py # M001 metrics
plots/
forecasting_leaderboard.py # leaderboard + boxplot figures
single_merchant_graphs.py # M001 actual-vs-forecast figure
figures/ # publication figures
requirements.txt
README.md
From the project root:
# Optional: create and activate a virtual environment
python -m venv .venv
# Windows
.venv\Scripts\activate
# macOS / Linux
# source .venv/bin/activate
pip install -r requirements.txt
# Regenerate synthetic raw data and dictionary (optional; committed raw file already exists)
python data/generator/data_generator.py
# Run the full portfolio model comparison
python source/all_merchants.py
# Generate portfolio figures (requires leaderboard.xlsx from the step above)
python plots/forecasting_leaderboard.py
# Optional: regenerate the illustrative single-merchant figure
python plots/single_merchant_graphs.pyNo cloud services, API keys, or environment variables are required.
- The dataset is synthetic; results may differ on real merchant businesses
- Merchants are heterogeneous; portfolio averages can hide merchant-level differences
- Forecast quality should be reviewed both at the portfolio level and by merchant
- This repository demonstrates a benchmarking framework; it is not a production forecasting service
- Results are associative evaluation outcomes, not causal claims about business interventions
- Model rankings should be refreshed as new months of data arrive
- The single-merchant example does not implement every portfolio model; use
source/all_merchants.pyfor the full nine-model comparison
- Python
- pandas
- numpy
- statsmodels
- scikit-learn
- matplotlib
- openpyxl