SPB Git

Repository Public

experiments.json 12.8 KB 1 lines json

1{"experiments":[{"id":"880c79ea-a091-43c4-bc0e-d38f5f81f624","workspaceId":"s-3d04acf2-235e-46e3-8c2d-bf26b0ba3c81","sessionId":"3d04acf2-235e-46e3-8c2d-bf26b0ba3c81","generationId":"0bc4b1ce-439f-451e-a893-f49987b53db6","seq":33,"kind":"experiment","title":"Stationarity battery (ADF, PP, DF-GLS, KPSS, Zivot-Andrews)","hypothesis":"Inflation, unemployment and T-bill are (non-)stationary; verdict depends on regime.","method":"1959Q2-2009Q3 (n=202, first obs dropped: placeholder 0). Full sample and splits 1959-83 / 1984-2009.","codeRef":null,"codeHash":null,"envHash":"af5ef3d79d245a52","runId":null,"status":"ok","metrics":{"infl ADF p full":0.027,"infl ADF p 1959-83":0.456,"infl ADF p 1984-09":0,"unemp ADF p full":0.111,"tbill ADF p full":0.28,"ZA p infl":0.233},"diagnostics":null,"artifacts":["artifacts/stationarity.csv"],"notes":"Inflation: ADF/PP/DF-GLS reject on full sample, but not ADF in 1959-83; KPSS rejects in both subsamples. Unemp and T-bill behave as near-unit-root. PP over-rejects given strong negative MA(1) in quarterly CPI inflation. Consistent with a mean shift/regime in inflation rather than a pure unit root.","error":null,"startedAt":1795213200593,"finishedAt":1795213200593,"durationMs":null},{"id":"28b5d2e8-2c04-44db-b96e-e32e35a0cb42","workspaceId":"s-3d04acf2-235e-46e3-8c2d-bf26b0ba3c81","sessionId":"3d04acf2-235e-46e3-8c2d-bf26b0ba3c81","generationId":"0bc4b1ce-439f-451e-a893-f49987b53db6","seq":34,"kind":"experiment","title":"Break detection with ruptures (Dynp/PELT/Binseg/BottomUp) on levels and multivariate","hypothesis":null,"method":"ruptures 1.1.10, l2 cost, min_size=16q, BIC for k<=6; PELT with penalty 2ln n and 4ln n.","codeRef":null,"codeHash":null,"envHash":"af5ef3d79d245a52","runId":null,"status":"ok","metrics":{"infl PELT":"1973Q1, 1981Q4","infl Dynp BIC k=5":"1965Q4,1973Q1,1977Q4,1981Q4,1991Q1","tbill PELT":"1966Q1,1978Q3,1982Q3,1991Q1,2001Q3","unemp PELT":"1974Q4,1987Q2"},"diagnostics":null,"artifacts":null,"notes":"Robust across algorithms: 1973Q1 and 1981Q4 for inflation. Mean-shift BIC overselects breaks on persistent series (unemp k=6, tbill k=5) - treat as descriptive. Multivariate BIC never reached a minimum for k<=6: unreliable, dropped.","error":null,"startedAt":1795213200594,"finishedAt":1795213200594,"durationMs":null},{"id":"95294c79-17d5-4d34-b387-95e51f13fbea","workspaceId":"s-3d04acf2-235e-46e3-8c2d-bf26b0ba3c81","sessionId":"3d04acf2-235e-46e3-8c2d-bf26b0ba3c81","generationId":"0bc4b1ce-439f-451e-a893-f49987b53db6","seq":35,"kind":"experiment","title":"Regression breaks: sup-Wald with recursive wild bootstrap, Bai-Perron via CostLinear, Chow, OLS-CUSUM","hypothesis":null,"method":"Phillips ADL(2): pi_t on pi_{t-1},pi_{t-2},u_{t-1}; Taylor partial adjustment: i_t on i_{t-1}, 4q mean inflation, u_t. 15% trimming, 499 wild bootstrap draws.","codeRef":null,"codeHash":null,"envHash":"af5ef3d79d245a52","runId":null,"status":"ok","metrics":{"PC supW all":44.4,"PC date":"1981Q4","PC p":0.004,"PC slope-only supW p":0.433,"TR supW all p":0.076,"TR date":"1980Q3","TR slope-only p":0.766,"OLS-CUSUM PC p":0.256,"OLS-CUSUM TR p":0.393},"diagnostics":null,"artifacts":["artifacts/pc_regimes.csv","artifacts/taylor_regimes.csv"],"notes":"Break in Phillips curve is in intercept/persistence (p=0.002), not in the unemployment slope alone (p=0.43). Taylor rule break only marginal (p=0.076); BP BIC suggests 1980Q4 and 1985Q1. OLS-CUSUM has no power here. Chow 1984Q1 uses asymptotic F with no heteroskedasticity robustness.","error":null,"startedAt":1795213200595,"finishedAt":1795213200595,"durationMs":null},{"id":"4339e692-e113-46b1-98f0-ec1c8078075d","workspaceId":"s-3d04acf2-235e-46e3-8c2d-bf26b0ba3c81","sessionId":"3d04acf2-235e-46e3-8c2d-bf26b0ba3c81","generationId":"0bc4b1ce-439f-451e-a893-f49987b53db6","seq":59,"kind":"experiment","title":"OOS forecasting horse race, expanding window (14 models, h=1,4,8)","hypothesis":"Phillips-curve / VAR / regime-switching models beat a univariate AR for US inflation out of sample.","method":"Pseudo-OOS, origins 1979Q4-2009Q2 (n=119/116/112), direct forecasts of mean annualised inflation over t+1..t+h. Models: RW1, RW4, AR(BIC<=4), IMA(1,1), UCSV particle filter, MS-AR(1), MS-AR(3), MS-PC (h=1), ADL Phillips (3 variants), VAR(3 vars, BIC), Ridge (40 predictors), equal-weight combination. Scripts in code/fc_*.py run in parallel via research_run.","codeRef":null,"codeHash":null,"envHash":"af5ef3d79d245a52","runId":null,"status":"ok","metrics":{"RMSFE AR h1/h4/h8":"2.595/1.823/1.629","RMSFE UCSV":"2.466/1.688/1.701","rel UCSV":"0.950/0.926/1.044","rel PC_u":"1.052/1.028/1.156","rel VAR3":"1.082/1.260/1.520","rel MS_AR3":"1.018/1.012/1.067","DM p UCSV vs AR":"0.176/0.254/0.734"},"diagnostics":{"look_ahead":"none: direct regressions use pairs with s+h<=t; UCSV is a filter; MS warm-started on past data only","tests":"DM-HLN (squared loss), Clark-West for nested","caveat":"final-vintage data, not real-time"},"artifacts":["artifacts/rmsfe_expanding.csv","artifacts/dm_vs_AR_expanding.csv","code/fc_linear.py","code/fc_ms.py","code/fc_ms3.py","code/fc_ucsv.py"],"notes":"No model beats AR significantly at 5%. UCSV is the best point forecaster at h=1,4. Economic models (Phillips, VAR) are worse, VAR/Ridge significantly at 10% only in the wrong direction.","error":null,"startedAt":1795213404597,"finishedAt":1795213404597,"durationMs":null},{"id":"aea472a2-15e7-4a04-8b29-d05371fc388c","workspaceId":"s-3d04acf2-235e-46e3-8c2d-bf26b0ba3c81","sessionId":"3d04acf2-235e-46e3-8c2d-bf26b0ba3c81","generationId":"0bc4b1ce-439f-451e-a893-f49987b53db6","seq":60,"kind":"experiment","title":"Robustness: rolling windows 60/40, MAE, start 1985, ex-GFC, subperiods, MCS, Phillips variants","hypothesis":null,"method":"Rolling 60 and 40 quarters; MAE loss; origins from 1984Q4; excluding targets overlapping 2008Q3+; Hansen-Lunde-Nason MCS (arch, block bootstrap, 10%); Phillips variants (delta u, backward-looking MA gap, real-rate proxy); MS-AR(3) via simulation.","codeRef":null,"codeHash":null,"envHash":"af5ef3d79d245a52","runId":null,"status":"ok","metrics":{"post-1985 rel UCSV h1/4/8":"0.929/0.843/0.821","post-1985 DM p UCSV h4":0.071,"MCS 1985+ h8":"Comb, IMA11, MS_AR3, RW4, Ridge, UCSV (AR, PC, VAR excluded)","MCS full sample":"all 11 models retained at every horizon","MAE rel UCSV":"0.967/0.950/0.938","Phillips variants rel AR h1":"1.020/1.015/1.049"},"diagnostics":null,"artifacts":["artifacts/robust_window_rmsfe.csv","artifacts/robust_mae_rel.csv","artifacts/robust_start1985_rel.csv","artifacts/subsample_results.csv","artifacts/mcs_results.csv","artifacts/dm_post1985.csv"],"notes":"Conclusions stable in sign: Phillips/VAR never beat AR. Anchored-mean models (UCSV, IMA, RW4) win at h=4,8 after 1990 and in rolling windows. MCS has little power in full sample. Many comparisons: nominal p-values not corrected for multiple testing. Rolling-window MS fits fail often (13 to 34 failures out of ~120-350), those origins left NaN and dropped on common samples.","error":null,"startedAt":1795213404599,"finishedAt":1795213404599,"durationMs":null},{"id":"82b13f27-3807-404d-bb88-ffa254ff9158","workspaceId":"s-3d04acf2-235e-46e3-8c2d-bf26b0ba3c81","sessionId":"3d04acf2-235e-46e3-8c2d-bf26b0ba3c81","generationId":"0bc4b1ce-439f-451e-a893-f49987b53db6","seq":61,"kind":"experiment","title":"Residual diagnostics (AR(3), Phillips ADL, Taylor, VAR(2), MS-AR(1))","hypothesis":null,"method":null,"codeRef":null,"codeHash":null,"envHash":"af5ef3d79d245a52","runId":null,"status":"ok","metrics":{"AR3 LB4 p":0.951,"AR3 ARCH4 p":0,"PC LB4 p":0.0038,"Taylor kurtosis":17.6,"VAR whiteness p":0,"VAR Granger unemp->infl p":0.976,"VAR Granger tbill->infl p":0.0064,"MS-AR std. resid ARCH4 p":0.688,"MS-AR std. resid LB4 p":0},"diagnostics":null,"artifacts":["artifacts/residual_diagnostics.csv"],"notes":"All linear models: strong ARCH and fat tails; MS-AR(1) removes ARCH but leaves serial correlation (AR(1) too short), which motivated the MS-AR(3) run. VAR(2) stable (max root modulus 0.955).","error":null,"startedAt":1795213404602,"finishedAt":1795213404602,"durationMs":null},{"id":"148b6706-4f19-4a52-ad01-b74ff1a0c262","workspaceId":"s-3d04acf2-235e-46e3-8c2d-bf26b0ba3c81","sessionId":"3d04acf2-235e-46e3-8c2d-bf26b0ba3c81","generationId":"0bc4b1ce-439f-451e-a893-f49987b53db6","seq":62,"kind":"experiment","title":"pip install ruptures via subprocess (failed attempt)","hypothesis":null,"method":null,"codeRef":null,"codeHash":null,"envHash":"af5ef3d79d245a52","runId":null,"status":"failed","metrics":null,"diagnostics":null,"artifacts":null,"notes":"pip through subprocess failed with SSL certificate error; install_package(\"ruptures\") worked (v1.1.10). No effect on results.","error":null,"startedAt":1795213404604,"finishedAt":1795213404604,"durationMs":null},{"id":"c046a30e-3425-4b1f-9277-c749ddb0361b","workspaceId":"s-3d04acf2-235e-46e3-8c2d-bf26b0ba3c81","sessionId":"3d04acf2-235e-46e3-8c2d-bf26b0ba3c81","generationId":"0bc4b1ce-439f-451e-a893-f49987b53db6","seq":63,"kind":"experiment","title":"Multivariate BIC break selection with ruptures (infl, unemp, tbill standardised)","hypothesis":null,"method":null,"codeRef":null,"codeHash":null,"envHash":"af5ef3d79d245a52","runId":null,"status":"abandoned","metrics":null,"diagnostics":null,"artifacts":null,"notes":"BIC decreased monotonically up to k=6 (never reached a minimum), so no break count could be chosen; dropped in favour of univariate PELT/Dynp and regression-based Bai-Perron. Also: first version of fc_ms.py read regime intercepts from the wrong parameter slots (p[:2] are transition probabilities); caught by inspecting param_names before any run and corrected.","error":null,"startedAt":1795213404606,"finishedAt":1795213404606,"durationMs":null},{"id":"0282c8be-39a8-4f89-b738-24a235be164e","workspaceId":"s-3d04acf2-235e-46e3-8c2d-bf26b0ba3c81","sessionId":"3d04acf2-235e-46e3-8c2d-bf26b0ba3c81","generationId":"0bc4b1ce-439f-451e-a893-f49987b53db6","seq":65,"kind":"experiment","title":"Rolling 40-quarter Phillips slope and Taylor long-run coefficient (HAC)","hypothesis":null,"method":"OLS on 40-quarter windows, HAC (3 lags) standard errors, delta method for the long-run Taylor coefficient. Significance = |t| > 1.96 per window, windows overlap (not independent tests).","codeRef":null,"codeHash":null,"envHash":"af5ef3d79d245a52","runId":null,"status":"ok","metrics":{"phillips_slope_min":-1.56,"phillips_slope_min_date":"1984Q2","phillips_slope_max":0.16,"phillips_slope_max_date":"2005Q3","share_windows_abs_t_gt_1.96":0.27,"taylor_share_windows_LR_gt_1":0.14,"taylor_share_CI_upper_below_1":0.36},"diagnostics":null,"artifacts":["artifacts/rolling_pc.csv","artifacts/rolling_taylor.csv"],"notes":"The 27% is a share of windows with |t|>1.96, not a p-value. Windows overlap heavily, so this describes instability and is not a formal test.","error":null,"startedAt":1795213484035,"finishedAt":1795213484035,"durationMs":null},{"id":"00d7f21b-1726-4464-b9df-f54d4c489186","workspaceId":"s-3d04acf2-235e-46e3-8c2d-bf26b0ba3c81","sessionId":"3d04acf2-235e-46e3-8c2d-bf26b0ba3c81","generationId":"c776e64f-f227-42c2-ae53-ca7ed1a1becc","seq":82,"kind":"experiment","title":"Test de fluctuation de Giacomini-Rossi (2010), UCSV vs AR, h=4","hypothesis":"La précision relative UCSV/AR à h=4 est-elle stable dans le temps ? H0 : espérance de la différence de perte quadratique nulle à chaque date.","method":"Statistique codée à la main (code/gr_fluct.py) sur les prévisions enregistrées (linear_expanding.csv, ucsv_g0.2.csv), P=116 origines 1979T4-2008T3. Variance de long terme Bartlett (L=8 retenu après contrôle de taille ; L=3,4 testés). Seuils par simulation du mouvement brownien (60 000 réplications, grille 2 320), comparés à la table de l'article rappelée de mémoire. Fenêtres m=29/35/46/58 ; pertes SE et AE ; UCSV gamma 0,1 et 0,2.","codeRef":"code/gr_fluct.py","codeHash":"0fdad7028ae27c45","envHash":"af5ef3d79d245a52","runId":"331159db-1fa0-4a55-a785-2663a77efcec","status":"ok","metrics":{"P":116,"rel_RMSFE":0.926,"DM_p":0.254,"maxF_m35_L8":2.385,"cv5_m35":3.07,"cv10_m35":2.809,"maxF_m46_L8":2.014,"cv5_m46":2.929,"rejets_sur_36_specs":0,"taille_5pct_L4_m35_iid":0.081,"taille_5pct_L8_m35_iid":0.05,"puissance_decalage_0.5sd":0.8,"RMSFE_sans_2005_09_rel":1.012,"DM_p_sans_2005_09":0.8},"diagnostics":{"validation_cv":"simulation ~0,05-0,09 au-dessus de la table de l'article (rappelée de mémoire, PDF non rouvert)","taille":"surdimensionné pour L=3,4 (8-15 %), nominal pour L=8","look_ahead":"aucun : prévisions déjà pseudo-hors-échantillon"},"artifacts":["artifacts/gr_results.csv","artifacts/gr_critical_values.csv","artifacts/gr_size_check.csv","artifacts/gr_power_check.csv","artifacts/figures/gr_fluctuation_ucsv_vs_ar_h4.png"],"notes":"Non-rejet partout ; le gain total de UCSV vient presque entièrement des cibles 2005-2009 (16 obs, 4 trimestres dominent). Sans elles rel RMSFE = 1,01.","error":null,"startedAt":1795213881141,"finishedAt":1795213881141,"durationMs":null}]}
Files main
Menu

Theme