Headline

The forecast model predicts flat revenue of $7,998/month for the next 6 months, but a 40.58% MAPE indicates poor accuracy and masks a severe 69.1% year-over-year revenue decline.

Purpose

This analysis applies an ETS(A,N,N) time series model to 13 months of historical e-commerce revenue data (4,882 transactions after cleaning) to forecast the next 6 months with confidence intervals. The objective is to project future revenue and identify seasonal patterns to support inventory and cash flow planning. However, the model's poor accuracy and the underlying business trend require careful interpretation.

Key Findings

MAPE (Mean Absolute Percentage Error): 40.58% — classified as poor accuracy. For demand forecasting, <20% is good; this model's predictions will miss actual values by roughly 2 in 5 cases.
6-Month Forecast: $7,998/month × 6 = $47,988 total — a flat projection with no growth (1.7% growth rate is negligible).
Year-over-Year Decline: −69.1% — revenue has collapsed compared to the prior year, indicating a fundamental business problem, not a forecasting issue.
Confidence Intervals (95%): $3,343 to $12,653 per month — extremely wide, reflecting high uncertainty. The lower bound is 58% below the point forecast.
Model Fit: MASE of 0.4 (better than naive baseline) and MAE of $2,778 suggest the model captures some structure, but the ETS(A,N,N) specification (additive error, no trend, no seasonality) is overly simplistic.
Data Quality: 97.6% retention after removing returns and missing dates; no detected seasonality strength, despite April ($14,231) and November ($13,671) showing 2.7× higher revenue than February ($5,326).

Interpretation

The model is technically sound but operationally unreliable. A 40.58% MAPE means forecasts are too volatile for tactical decisions (staffing, procurement). The flat forecast masks the real story: revenue has dropped 69% year-over-year, yet the model projects stability. This suggests either a structural break in the business (market loss, product discontinuation, operational disruption) or data quality issues (the dataset spans Dec 2010–Dec 2011; if this is historical data, the forecast is retrospective). The wide 95% CI ($3,343–$12,653) reflects the model's inability to pin down a reliable trend. Seasonal patterns exist (April peak, February trough) but were not captured by the ETS(A,N,N) specification.

Context

The analysis uses only 13 months of history, which is minimal for robust forecasting. The 69.1% YoY decline is the dominant signal and should be investigated before relying on any forecast. If this is real business data, the forecast is only useful as a lower-bound estimate; actual revenue could be higher if the decline has stabilized, or lower if it continues. The model's MASE < 1 indicates it outperforms a naive (no-change) baseline, but that baseline is itself unreliable given the sharp YoY drop.