Allisterh/food-inflation-at-risk-nigeria

Replication package for "Level, Not Shape: Why Conditional Tail Forecasts Fail at Structural Breaks in Granular Food Prices" — 90th-percentile food inflation-at-risk forecasts across Nigeria's 2023 subsidy and exchange-rate break, on a 139,917-observation state–commodity price panel. # Level, Not Shape — replication package Manuscript prepared for submission to *Journal of Forecasting*. ## What changed relative to the previous draft The earlier draft ("Food Inflation-at-Risk under Delayed Feedback") proposed a regime-aware meta-controller and reported, honestly, that it could not be shown superior to the best individual expert, that it failed during recovery, and that the evaluation was exploratory rather than confirmatory. That framing concedes the paper's own headline claim, and the evaluation ran on 28 national series rather than the full state-commodity panel. This version keeps the data and the delayed-feedback discipline but changes the claim to one the evidence actually supports: | | Previous draft | This version | |---|---|---| | Primary panel | 28 national item series, 638 forecasts | 37 states x 43 items, 58,864 forecasts | | Evaluation origins | 23 | 37 (26 stable, 11 break) | | Hyperparameter tuning | selected on 2022 | none — all constants fixed *a priori* | | Train/test boundary | chronological | chronological **and purged** | | Benchmarks | five self-constructed variants | + quantile regression, gradient boosting, CAViaR, adaptive conformal | | Inference | descriptive month-block bootstrap | Diebold-Mariano (HAC + HLN), Giacomini-White, Hansen MCS, Kupiec, Christoffersen, dynamic quantile | | Headline claim | "our controller is promising" | "tail failure at breaks is a level effect, not a conditioning effect" | | Inferential status | exploratory | confirmatory out-of-sample | ## Main findings 1. Conditional quantile models fail **worse** than unconditional ones at the break: gradient-boosting coverage 25.8% vs 44.2% for a fixed empirical quantile, against a 90% nominal target (DM *p* < 0.001; both conditional models eliminated from the 90% model confidence set). 2. The failure is a level effect. A single scalar recovers 88.1% of recoverable loss during the break. After level correction the three forecasts are statistically …