feat: publish verified market comparison with robust accounting

.gitattributes

@@ -0,0 +1,2 @@
+# Frozen evidence is hashed as bytes; checkouts must not rewrite line endings.
+docs/experiments/real-sp500/** -text

.github/workflows/ci.yml

@@ -42,6 +42,9 @@ jobs:
       - name: Docs integrity (mkdocs nav resolves)
         run: python scripts/check_docs.py
 
+      - name: Independently verify the saved real-market comparison
+        run: python scripts/verify_comparison.py docs/experiments/real-sp500
+
       - name: Unit & regression tests (fast)
         run: pytest -m "not slow"
 

.gitignore

@@ -56,3 +56,6 @@ data/processed/*
 .pytest_tmp/
 # MkDocs build output
 site/
+outputs/
+research/real_runs/
+research/comparison/

README.md

@@ -1,5 +1,10 @@
 # AlphaForge
 
+[Real-market comparison](docs/research-comparison.md): frozen data, shared
+walk-forward coverage, equal-weight and momentum baselines, three model seeds,
+one explicit size-feature ablation, and 22 base/doubled-cost cases. Source
+history and availability limitations keep this experiment exploratory.
+
 **Python toolkit for factor research, portfolio construction, and backtesting**
 
 [![CI](https://github.com/dev-belly/alphaforge/actions/workflows/ci.yml/badge.svg)](https://github.com/dev-belly/alphaforge/actions/workflows/ci.yml)

configs/real_sp500.yaml

@@ -23,7 +23,7 @@ data:
   provider: local
   universe: SP500
   start_date: "2000-01-01"
-  end_date: "2019-12-31"
+  end_date: "2019-11-14"
   cache_dir: data/processed
   raw_dir: data/raw
   storage_backend: parquet
@@ -88,7 +88,7 @@ risk:
 
 backtest:
   start_date: "2005-01-01"
-  end_date: "2019-12-31"
+  end_date: "2019-11-14"
   rebalance: monthly
   initial_capital: 10_000_000
   execution_lag_days: 1
@@ -103,7 +103,7 @@ cost:
   min_commission: 0.0
 
 reporting:
-  output_dir: research/reports
+  output_dir: research/real_runs/report
   figure_dpi: 130
   figure_format: png
 

docs/experiments/real-sp500/folds.csv

@@ -0,0 +1,55 @@
+method,seed,fold,rank_ic_mean,rank_ic_std,n_days,icir
+ridge,11,0,0.04412592633746016,0.0966486106764495,252,0.4565603791779325
+ridge,11,1,0.018713655037396176,0.08398618101527376,252,0.22281826380452877
+ridge,11,2,-0.02324649092185162,0.0944463490283868,252,-0.24613435205276862
+ridge,11,3,0.02219967924672747,0.08837599819123425,251,0.2511957963822964
+ridge,11,4,-0.005444657308135957,0.12453645599247631,251,-0.04371938533777522
+ridge,11,5,0.009387714645803524,0.10163328992091236,200,0.09236850104044383
+ridge,29,0,0.04412592633746016,0.0966486106764495,252,0.4565603791779325
+ridge,29,1,0.018713655037396176,0.08398618101527376,252,0.22281826380452877
+ridge,29,2,-0.02324649092185162,0.0944463490283868,252,-0.24613435205276862
+ridge,29,3,0.02219967924672747,0.08837599819123425,251,0.2511957963822964
+ridge,29,4,-0.005444657308135957,0.12453645599247631,251,-0.04371938533777522
+ridge,29,5,0.009387714645803524,0.10163328992091236,200,0.09236850104044383
+ridge,42,0,0.04412592633746016,0.0966486106764495,252,0.4565603791779325
+ridge,42,1,0.018713655037396176,0.08398618101527376,252,0.22281826380452877
+ridge,42,2,-0.02324649092185162,0.0944463490283868,252,-0.24613435205276862
+ridge,42,3,0.02219967924672747,0.08837599819123425,251,0.2511957963822964
+ridge,42,4,-0.005444657308135957,0.12453645599247631,251,-0.04371938533777522
+ridge,42,5,0.009387714645803524,0.10163328992091236,200,0.09236850104044383
+lightgbm,11,0,0.022477486143247915,0.08734886863632098,252,0.25733002034443564
+lightgbm,11,1,0.002234562428456477,0.05797573874961744,252,0.038543060884605324
+lightgbm,11,2,-0.0011553035855261433,0.10126092459062845,252,-0.011409174765061003
+lightgbm,11,3,0.03166639744953009,0.06502624933653303,251,0.48697868587876997
+lightgbm,11,4,0.031668053472560784,0.08124079588245281,251,0.3898048157773004
+lightgbm,11,5,0.03433905815716499,0.055633472983191,200,0.6172373629728299
+lightgbm,29,0,0.025787883052372746,0.08480912689437926,252,0.3040696679319527
+lightgbm,29,1,-0.0030296977978654153,0.057293289941090675,252,-0.05288049963583118
+lightgbm,29,2,0.00907244205085725,0.09838261707558023,252,0.09221590480651223
+lightgbm,29,3,0.032402329646731974,0.06749892297971691,251,0.48004217276872274
+lightgbm,29,4,0.017454562276111394,0.08673166092103583,251,0.20124787293077168
+lightgbm,29,5,0.039869933952373565,0.06175223753608078,200,0.6456435514434314
+lightgbm,42,0,0.020528439407260922,0.09214009289193455,252,0.22279594867933838
+lightgbm,42,1,-0.0035977333884247465,0.0583879427725417,252,-0.06161774533554323
+lightgbm,42,2,0.005875082979020472,0.1014245680706104,252,0.05792563962343245
+lightgbm,42,3,0.03195550150870484,0.06497403962210874,251,0.4918195281463049
+lightgbm,42,4,0.017572815275974017,0.08240420495439846,251,0.21325143887619097
+lightgbm,42,5,0.04023812238863307,0.05787186037686153,200,0.6952968528504602
+ridge_without_size_feature,11,0,0.04359069019775172,0.09661460929879445,252,0.4511811465587083
+ridge_without_size_feature,11,1,0.016504867360824224,0.08288305728509346,252,0.19913439370428038
+ridge_without_size_feature,11,2,-0.02476588151153135,0.09478300955559105,252,-0.2612903053791086
+ridge_without_size_feature,11,3,0.02027092225650099,0.09210790690979509,251,0.22007798175625798
+ridge_without_size_feature,11,4,-0.007838736870762348,0.12484721777420321,251,-0.0627866364226023
+ridge_without_size_feature,11,5,0.0029383498668230815,0.10368188633286714,200,0.02834005023201073
+ridge_without_size_feature,29,0,0.04359069019775172,0.09661460929879445,252,0.4511811465587083
+ridge_without_size_feature,29,1,0.016504867360824224,0.08288305728509346,252,0.19913439370428038
+ridge_without_size_feature,29,2,-0.02476588151153135,0.09478300955559105,252,-0.2612903053791086
+ridge_without_size_feature,29,3,0.02027092225650099,0.09210790690979509,251,0.22007798175625798
+ridge_without_size_feature,29,4,-0.007838736870762348,0.12484721777420321,251,-0.0627866364226023
+ridge_without_size_feature,29,5,0.0029383498668230815,0.10368188633286714,200,0.02834005023201073
+ridge_without_size_feature,42,0,0.04359069019775172,0.09661460929879445,252,0.4511811465587083
+ridge_without_size_feature,42,1,0.016504867360824224,0.08288305728509346,252,0.19913439370428038
+ridge_without_size_feature,42,2,-0.02476588151153135,0.09478300955559105,252,-0.2612903053791086
+ridge_without_size_feature,42,3,0.02027092225650099,0.09210790690979509,251,0.22007798175625798
+ridge_without_size_feature,42,4,-0.007838736870762348,0.12484721777420321,251,-0.0627866364226023
+ridge_without_size_feature,42,5,0.0029383498668230815,0.10368188633286714,200,0.02834005023201073

docs/experiments/real-sp500/index.html

@@ -0,0 +1,10 @@
+<!doctype html><html lang="en"><meta charset="utf-8">
+<meta name="viewport" content="width=device-width,initial-scale=1"><title>AlphaForge · Frozen comparison</title>
+<style>:root{font-family:system-ui;color-scheme:dark;background:#111319;color:#eee}main{max-width:1100px;margin:auto;padding:2rem}h1{font-size:3rem}a{color:#d7fa73}table{border-collapse:collapse;width:100%}td,th{padding:.6rem;text-align:left;border-bottom:1px solid #444;white-space:nowrap}.scroll{overflow:auto}aside{padding:1rem;border-left:3px solid #ffb994}</style>
+<main><p>EXPLORATORY_REAL_MARKET</p><h1>One dataset.<br>Declared comparisons.</h1>
+<p>840,736 samples · 24 usable features · 551,698 shared forecasts.
+Models, the fixed momentum baseline and equal-weight baseline share dates, outcomes, execution lag and cost rules.</p>
+<aside>Seed variation is not market sampling uncertainty or proof of tradable alpha.<br>Cost inputs are fixed scenarios, not measured execution costs.<br>The equal-weight baseline differs in diversification; its holdings rule is explicit.<br>Source hashes establish identity, not complete historical data quality.<br>Historical membership is incomplete; survivorship/selection bias remains.<br>Financial availability is an assumed lag, not actual publication history.<br>Size uses a dollar-volume proxy, not measured market capitalization.</aside>
+<h2>All methods and cost scenarios</h2><div class="scroll"><table><tr><th>Method</th><th>Costs</th><th>Runs</th><th>Net CAGR</th><th>Net Sharpe mean ± seed std</th><th>Max drawdown</th></tr><tr><td>equal_weight</td><td>1×</td><td>1</td><td>12.75%</td><td>1.141 ± 0.000</td><td>-18.70%</td></tr><tr><td>equal_weight</td><td>2×</td><td>1</td><td>12.70%</td><td>1.137 ± 0.000</td><td>-18.71%</td></tr><tr><td>lightgbm</td><td>1×</td><td>3</td><td>18.23%</td><td>1.337 ± 0.022</td><td>-21.46%</td></tr><tr><td>lightgbm</td><td>2×</td><td>3</td><td>17.51%</td><td>1.290 ± 0.023</td><td>-21.64%</td></tr><tr><td>ridge</td><td>1×</td><td>3</td><td>12.85%</td><td>1.005 ± 0.000</td><td>-22.06%</td></tr><tr><td>ridge</td><td>2×</td><td>3</td><td>12.20%</td><td>0.960 ± 0.000</td><td>-22.20%</td></tr><tr><td>ridge_without_size_feature</td><td>1×</td><td>3</td><td>12.24%</td><td>0.955 ± 0.000</td><td>-21.68%</td></tr><tr><td>ridge_without_size_feature</td><td>2×</td><td>3</td><td>11.58%</td><td>0.909 ± 0.000</td><td>-21.84%</td></tr><tr><td>single_momentum</td><td>1×</td><td>1</td><td>11.38%</td><td>0.871 ± 0.000</td><td>-23.40%</td></tr><tr><td>single_momentum</td><td>2×</td><td>1</td><td>10.95%</td><td>0.842 ± 0.000</td><td>-23.47%</td></tr></table></div>
+<p><a href="summary.json">Source identity, exact protocol and all runs</a> · <a href="metrics.csv">Metrics CSV</a> · <a href="folds.csv">Fold results</a> · <a href="daily_returns.csv.gz">Daily returns (gzip CSV)</a> · <a href="manifest.json">Hashes</a></p>
+<p>This is a protocol comparison. Its real-market variant remains exploratory until the declared source limitations are resolved. No test result selects a model or seed for deployment.</p></main></html>

docs/experiments/real-sp500/manifest.json

@@ -0,0 +1,10 @@
+{
+  "dataset_sha256": "8906ace6a84d22bdc78a05fedcd75498d964dd56555da00492892f5eb7a9a7cd",
+  "files_sha256": {
+    "daily_returns.csv.gz": "87eb8aeee44cd396844433d06cc0b3d937041c9710d55e1418acebc2d6b09ede",
+    "folds.csv": "8ae472f7117df46258db54cfff3a5bb349df99e04fdeceb1e3ce37a7c8ac88df",
+    "index.html": "aa9003c9a31717baafba4838c7955ed62ec63ca76e424a518af4cb6e8b44ac2e",
+    "metrics.csv": "41b0d117631aa0613d0d054b817a959c3792079ec775ea3c73a409265d89d106",
+    "summary.json": "8f51affb4f487f9cb25995e89de91d831b5e2e4933ea69495478d4883066a2fb"
+  }
+}