Extreme precipitation events represent high-impact hazards, motivating the development of lightweight early-warning models that can be trained directly from station observations. This paper investigates whether rolling Benford deviation features, widely used in forensic analytics and anomaly detection, provide additional predictive value for short-horizon forecasting of precipitation extremes. Using daily NOAA station data from five locations representing arid, semi-arid, and humid precipitation regimes, we formulate the task as predicting whether a station-specific extreme threshold ($P95$, estimated on the training period) will be exceeded within the next three days. A strong temporal baseline is constructed using lagged predictors, rolling statistics, seasonal encoding, and temperature variables when available. Benford features are computed in a past-only rolling manner across multiple window sizes (90, 180, and 365 days), including first-digit deviation (MAD), divergence measures, rounding ratios, and valid-sample counts. Evaluation using leakage-aware rolling-origin cross-validation shows that the baseline achieves strong performance (PR-AUC $\approx 0.76$--$0.84$), while Benford features provide no improvement: MAD-only features are nearly neutral, whereas richer Benford feature sets slightly reduce performance. The results further indicate that Benford reliability strongly depends on wet-day coverage, becoming unstable in arid climates and largely redundant in humid regimes. These findings suggest that Benford analysis is more suitable for meteorological data quality control and regime characterization than for short-term extreme-event prediction.