Cold §22 forecast measured ~215-233s on prod: §9.x layers re-execute the same
horizon/segment-invariant DB loads with identical args hundreds of times per
report (profiled: get_competitors x69, market_metrics x124, get_monthly_macro
x290). Add a per-report ContextVar cache (forecast_cache(), opened once in the
orchestrator) + @cached(key_builder) on the expensive §9.x loaders so each
unique load runs ONCE and reuses the same frozen, read-only instance.
Output is byte-identical (memoized producers are frozen dataclasses / read-only
Pydantic, callers never mutate; cache is per-report, discarded on exit; no-op
outside the report build). No concurrency, no signature changes.
- forecast_request_cache.py: ContextVar cache + cached() decorator (no-op
outside context, reentrant, _MISS sentinel for cached None)
- @cached on competitors/future_supply/market_metrics/macro_series/
sales_series/macro_coefficient/demand_normalization/regression loaders
- orchestrator: wrap build_site_finder_report in forecast_cache()
- 58 tests: key discrimination (call-counting regression guard), no-op-outside,
per-report isolation, reentrancy, frozen-producer canary, amplification proof
(real get_monthly_macro xN->1)
code-reviewer APPROVE (keys correct, mutation-safe, output identical). 1265
forecast/cache tests green. No new deps. Refs #1129.
Backtest (OOS directional hit-rate): single-best-lag compute_rate_sensitivity
is directionally noise (0.148 Source B EKB-wide, lag-unstable); the Almon
distributed-lag estimator (compute_district_rate_regression) is strictly less
noisy on every tier (0.407 Source B / 0.60 survivorship-free Source A,
lag-stable). Add a thin adapter compute_rate_regime_sensitivity mapping
DistributedLagFit onto the existing RateSensitivity contract (beta=long-run
sum-beta, confidence regression->medium / fallback->low, district=None->low and
no call) and repoint the three consumers (demand_normalization, product_scoring,
demand_supply_forecast). Magnitude bounded by the existing [0.5,1.2] clamp.
Reversible; compute_rate_sensitivity kept for the backtest. Consumer tests
repointed to the real Almon path (mutation-verified genuine) + adapter unit
tests + end-to-end fallback degradation. Forecasting suite 840 passed; ruff clean.
REOPENED 951-B §9.6.
PART A: fix look-ahead leakage in backtest_rate_sensitivity --detrend. The
ln(units) trend was fit over train+test then split, so test data shaped the
detrend and inflated the OOS hit-rate. _detrend_log now takes fit_n; backtest_tier
fits the trend on TRAIN months only (same split evaluate_oos uses) and projects
(a,b) point-in-time onto test. Default fit_n=None preserves prior behaviour.
PART B (DoD): new app/services/forecasting/regression.py — Almon polynomial
distributed-lag (deg 2) of Δln(district demand) on Δkey_rate lags 0..6 via
OLS-on-Almon-regressors (numpy lstsq) + per-lag reconstruction + manual
Newey-West HAC SEs (NO statsmodels). Output {best_lag_months, coef=long-run
multiplier, x_pct, r2, n, per_lag_coef, hac_se,...}; gate mirrors _elasticity_coef
(n<30 OR R²<0.1 OR Σβ≥0 → fallback); §9.6 phrase from the lag shape. ADVISORY,
shipped standalone (integration point documented), NOT wired — protects the live
compute_rate_sensitivity consumers.
125+31 tests (synthetic known-lag recovery, HAC computed/differs-from-OLS,
fallback gating, no-leakage detrend). ruff clean. Refs #978
