Controlled Feature Ablation for Inflation Nowcasting: When Is a Measured Gain Real?

This study evaluates whether a single free text sentiment feature improves one-month-ahead US CPI inflation nowcasts beyond six standard macroeconomic indicators, using a controlled ablation on a fixed ridge regression architecture with strict lag based leakage control. Over a 36 month walk forward test ($n=36$), removing the text feature raises mean absolute error from $0.120$ to $0.126$ percentage points, a $4.5\%$ degradation, and lowers the directional hit ratio from $63.9\%$ to $61.1\%$, while RMSE is unchanged ($0.197$ vs. $0.197$). A Diebold Mariano test with the Harvey-Leybourne-Newbold small sample correction on the absolute-error differential gives $p=0.11$. The sign of the gain persists across 24-, 36-, and 48-month horizons ($3.5\%$-$4.5\%$) but never reaches significance ($p=0.06$-$0.23$). Both learned models outperform non-learned baselines (naive month-over-month, 12-month rolling mean) by $22\%$. The measured improvement is found to be directionally consistent but statistically indistinguishable from noise at this sample size, suggesting that ablation claims for proprietary alternative data features should be held to the same significance bar rather than reported as unqualified gains.