# Decision log — WP23 | # | Decision | Rationale / evidence | |---|---|---| | D1 | WP22 used read-only; Task B predictors = WP22 real-time vintages (smoothed mode at each publication date), target = WP22 final smoothed index change | protocol §7.3; WP22 vintage database records both modes | | D2 | Task A training pairs carry a random horizon h ∈ {1,3,6,12} with market/neighbour features dated h months before the sale | otherwise horizon dependence cannot be learned (first HSHN run failed: untrained horizon embeddings, MAE 1–4) | | D3 | Tree models (LightGBM/XGBoost) and PyTorch models run in separate processes (`--family`) | libomp conflict on macOS: segmentation fault at backward or deadlock in the same process | | D4 | Quarterly origins for Task A (12), monthly origins with quarterly re-estimation for Task B (34–45 per horizon) | cost; 12-month targets must exist (last origin Sept. 2025 for h=12) | | D5 | No hyper-parameter tuning on test origins; fixed a priori (Table tab_hyper) | protocol §7.3 | | D6 | Giacomini–White reported only where ≥ 12 origins (Task B); DM with HAC(h−1) + HLN; MCS with T_max and stationary bootstrap on per-origin mean losses | short samples; cross-sectional dependence removed by averaging within origins | | D7 | QREFA = NNLS stack of QANSS, BHPF, DSGFN, RAREF (LightGBM excluded from the stack because of D3) | process isolation | | D8 | Simulation forecasters run with 10 replications (6 for the neural family), horizons 1/3/6 | compute on the node (RAREF HMM and torch models are slow) | | D9 | BHPF is conjugate/EM rather than MCMC; VI/HMC not available (PyMC absent) and unnecessary for a Gaussian hierarchy | exactness | | D10 | CMA boundaries unavailable; metropolitan comparisons not needed for forecasting tasks | inherited from WP22 |