Thales Research Notebook

Dated notes in this folder: one file per day, named YYYY-MM-DD.md. When multiple things happen in a day they all land in the same file, each entry headed with ## HH:MM — title. When a new day starts, create a new file.

• Current State (below) — updated in place; always reflects what is actually running in production.
• Daily notes — YYYY-MM-DD.md; append-only within a day.
• Index — at the bottom of this file, newest day first.

Earlier history not captured here is in git log and results/diagnostics/.

Current State — last updated 2026-04-06

Universe & Data

• Universe: Russell 1000 (universe.source: russell1000), min market cap $2B, min avg volume 500k
• Price data: Tiingo primary, yfinance fallback; 904 symbols loaded, 4.17M rows, range 2005-01-01 → ~2026-02-27 (stale)
• Factors: Fama-French 5 + momentum, data/factors/ff5_daily.parquet
• Macro: VIX, yield curve (T10Y2Y), unemployment claims (ICSA) in data/macro/
• Fundamentals: SEC EDGAR PiT (free), plus FMP for extended metrics. Used in meta-labeler features.
• Survivorship-free: opt-in via --survivorship-free flag; Wikipedia-scraped S&P 500 membership history in data/constituents/

Production Strategy

• Class: MomentumStrategy (src/thales/strategy/momentum.py)
• Signal: 12-2 momentum (lookback: 252, skip: 5) blended with smoothness (smoothness_window: 21, smoothness_weight: 0.5)
• Top-N: 50 stocks
• Weighting: HRP (strategy.construction.weighting: hrp) — crisis insurance, see 2026-04-06 HRP ablation
• Sizing: Half-Kelly (risk.sizing: half_kelly, kelly.fraction: 0.5, rolling_window_months: 24, min_observations: 8)

Portfolio Risk Controls

• Max single position: 10%
• Max sector: 35% (raised from 25% for momentum concentration)
• Target vol: 12% annualized (vol_target: 0.12, buffer 1%)
• Max leverage: 3.0x
• Drawdown kill-switch: 20% (drawdown_kill_switch: 0.20)
• Kill-switch re-entry: dispersion-regime aware (kill_switch_reentry: dispersion), min 21 cash days
• Dynamic vol target: off (dynamic_vol_target: false)

Execution

• Broker: Alpaca paper (account [REDACTED:acct], $10K)
• Schedule: GitHub Actions cron Mon-Fri 14:35 UTC via scripts/run_daily.py → thales run
• Mode: Daily signal diff (rebalance.mode: daily) with Tuesday alpha rebalance (rebalance_day: tuesday)
• Turnover cap: 5% daily (daily_turnover_cap: 0.05)
• Hysteresis: entry_band 0, exit_band 10 (hold until rank drops below N+10)
• Emergency exit: bypass turnover cap at rank 200+ or after 3 zero-volume days
• Execution timing: T+1 open fills (execution_timing: next_open) — models overnight gap + intraday returns
• TCA: 3-component implementation shortfall → data/processed/tca_log.jsonl
• Costs (backtest): flat 10 bps round-trip (costs.model: flat); impact model wired but disabled

ML Application — nothing live in production

1. Meta-labeler (src/thales/portfolio/meta_labeler.py) — disabled

• Config: meta_labeling.enabled: false
• Trained artifact: results/model/meta_labeler.pkl (last trained 2026-03-31, 3040 samples, in-sample Sharpe 0.87)
• Features (10): xs_fracdiff_z, xs_vol_z, rvol_5_63, xs_adf_z, cs_dispersion, spy_ma200_dist, vix, xs_ebit_yield_z, xs_sue_z, xs_accruals_z — has_fundamentals: true
• Pipeline: triple-barrier labels (AFML Ch 3), uniqueness weights (Ch 4), sequential bootstrap, ADF structural break filter, MDA importance, most recently PCA orthogonalization (Ch 8.4.2)
• Hyperparams: RF 200 trees, max_depth 4, barrier_mult 2.5, vol_span 20, label_horizon 42, purge_gap 42, embargo 5
• Sizing mode: cross_sectional (options: binary, continuous, relative_cdf, cross_sectional)
• CPCV history: PBO 100% with OHLCV only → 100% with AFML v2 → 100% with SEC fundamentals. The PCA orthogonalization (commit fbf0f37) is the current frontier — never CPCV'd, represents the most plausible path to a breakthrough.
• Stale warning: inline comment in settings.yaml still says "needs PiT fundamentals" but fundamentals are already wired.

2. LambdaRank (research/paused/ranker.py, RankerStrategy) — built, paused, archived 2026-05-26

• Learning-to-rank gradient boosting as the primary stock-selection signal (not a meta-label gate)
• 6-path CPCV: PBO 33.3% — first ML model below baseline (commit 090fb7a)
• 15-path CPCV: PBO 50% — matches baseline, does not beat it (commit ee3f00e)
• Archive note: moved from src/thales/strategy/ranker.py to research/paused/ranker.py on 2026-05-26 as part of the organization audit. Removed from cli.py strategy registry. To revive: move back into src/thales/strategy/, restore CLI registry entry, add tests, run through SHIP gate.

Supporting ML / Research Infrastructure

• Triple-barrier labeler: src/thales/preprocessing/labeler.py (deleted in cleanup 4e4c481; re-implementations live inside meta_labeler.py)
• Per-event features + Corwin-Schultz spread: src/thales/preprocessing/features.py (deleted in cleanup 4e4c481)
• Model store: src/thales/portfolio/model_store.py (save/load trained RF)
• Feature engineering: src/thales/features/indicators.py (momentum variants, fracdiff, RVOL, dispersion, volume confirmation, FF3 residual momentum)
• Fundamentals downloader: src/thales/data/fundamentals.py (SEC EDGAR, free, PiT-aligned)

Performance Snapshot (latest CPCV, 2026-04-06, HRP on, full 2005-2026 history)

Built but Disabled / Paused

Open Questions / Next Experiments

1. CPCV the PCA-orthogonalized meta-labeler (commit fbf0f37 is the untested frontier — could break PBO 100% ceiling)
2. LambdaRank tuning — feature set expansion, loss function tweaks to push PBO below 50%
3. Refresh price data (thales fetch) — currently stale at Feb 27
4. Validate dispersion regime filter in production config — currently off
5. TWAP order slicing when AUM reaches ~$250K+

Infrastructure

• Python: 3.13 venv
• Tests: 404 (CLAUDE.md count; includes 39 production integration tests in tests/test_execution/test_production_integration.py)
• Parallel CPCV: thales cpcv --workers N — 6 workers on M5 gives ~4 min wall time for 15 paths (~7.5× speedup)

Daily Log Index

• 2026-04-06 — HRP ablation (walk-forward vs CPCV disagree, HRP stays on); parallel CPCV validation; research notebook initialized