Headline

The forecast model predicts stable daily support ticket volume at 6.8 tickets per day, but with a MAPE of 44.55% — accuracy is too poor for reliable staffing decisions.

Purpose

This analysis forecasts daily support ticket volume over a 2-year historical period (730 days) and projects 30 days forward to enable staffing planning. The model decomposes observed volume into trend, seasonality, and noise components, then generates point forecasts with confidence intervals. Understanding forecast accuracy and the drivers of ticket volume is critical for allocating support resources efficiently.

Key Findings

Historical Volume: Average 6.8 tickets per day (range 1–17), with minimal trend change of −2% over the period — the volume is essentially flat.
Forecast Accuracy (MAPE): 44.55% — this is poor by forecasting standards (acceptable is <30%). The model cannot predict daily variation reliably.
Model Structure: ARIMA(0,0,0) with no autoregressive, differencing, or moving-average components — the model defaults to predicting the mean (6.8) every day.
Residual Diagnostics: Ljung-Box p-value of 0.694 (>0.05) indicates residuals are not autocorrelated — good. However, Shapiro-Wilk p-value of 0 shows residuals are non-normal, suggesting the model misses important patterns.
Day-of-Week Pattern: Thursday is busiest (+3.7% above average, 7.1 tickets), Sunday quietest (−5.1%, 6.5 tickets) — a small but consistent weekly rhythm.
30-Day Forecast: Point estimate of 7 tickets/day with 95% confidence interval [1, 12] — the wide bounds reflect high uncertainty.

Interpretation

The flat trend and low MAPE indicate that daily ticket volume is driven primarily by random variation rather than predictable patterns. The ARIMA(0,0,0) model essentially gives up and forecasts the historical mean. While the day-of-week effect is statistically real, it explains only ~3.7% of variation — not enough to improve staffing materially. The non-normal residuals suggest occasional spikes (up to 17 tickets) that the model cannot anticipate. For staffing, this means you cannot reliably predict tomorrow's load; you must plan for the range [1, 12] or use a fixed baseline of 7 with buffer capacity.

Context

The dataset spans 2 years (730 days) with no missing values, providing solid historical coverage. However, the poor forecast accuracy and lack of autocorrelation suggest that ticket volume may be driven by external factors (customer behavior, product issues, campaigns) not captured in the time series alone. Incorporating external regressors (e.g., product releases, marketing campaigns, customer count) could improve the model substantially.