Daily Stock Guide — research repository

Research, backtests, and infrastructure for the dailystockguide.com deployed strategy (currently v5 — a Chronos-filtered GBM ranking strategy on a point-in-time S&P 500 universe). This repo is the data layer, ML training, validation gauntlets, web frontend, and assorted experiments.

Quick links

Honest performance numbers

The deployed v5 strategy (as of 2026-05-17) is E2 ("LEAD-CAGR") (v5_pit_sp500_E2_win1_rcdadaptk_5050_chronos_p70_k2_invvol_cap0.4_minhold6_scoredrift) — a 50/50 portfolio of two sleeves of the same GBM+Chronos alpha (Chronos p70 filter, inverse-vol cap 0.4, rule-based rebalance min-6m/score-drift). Sleeve A = WIN1 (select=ml_3plus6, trigger=blend, the unchanged single-sleeve winner). Sleeve B = RC D + adaptive-breadth (trigger=ml_3plus6, select = a consensus/ml blend whose weight is regime-conditional — momentum-lean in a confirmed SPY bull, consensus-stable in normal/recovery — with conviction-adaptive basket breadth that holds 2 names on high conviction and widens to 3 when scores are bunched). The sleeves rebalance on different dates so their idiosyncratic variance decorrelates (E1's free-consistency lever) while Sleeve B's regime-timed blend + adaptive breadth add forward return and year-to-year consistency — orthogonal levers that stack. Combined live book is up to ~6 names. Validated on the augmented PIT panel with the canonical production sim — see IMPROVE_PICK_RCD_FINDINGS.md and IMPROVE_PICK_RCE1_FINDINGS.md.

Validation gauntlet (augmented PIT 2003–2025, 10 bps) that justified deploying E2 over the prior single-deploy E1 (WIN1+WIN2 50/50). The deployed live track now extends through 2026-04 — the backtest history is byte-identical (frozen, append-only), so these validation numbers stand; the current full-window figure is ~55% CAGR with the four 2026 live months added:

E2 is a strict Pareto improvement on E1: +4.7pp CAGR, higher Sharpe, better worst-rolling-5y DCA, and a +4.8pp higher forward (ex-2003–09) CAGR — at E1's identical −56% drawdown, WF 10/10, all four eras beating S&P-DCA, and 100% 10-year DCA-win. It is cost-insensitive, sits on a wide 50/50 mix-weight plateau (0.3–0.7), is more delisting-robust than E1, and has the strongest truly-out-of-sample holdout of any variant (untouched 2013–2026 Sharpe ~1.24 vs E1's 1.08). Honest caveat: the −56% drawdown is the 2008 GFC systemic event — a stock-picking lever cannot move it. E2 raises return and consistency at E1's drawdown; it does not make the strategy low-risk. The edge is still front-loaded in 2003–2009 and the interim drawdowns are still deep — narrowed, not removed.

The PIT correction adds 161 acquired/renamed large-caps (AGN, ANTM, ABMD, CELG, ATVI, AET, …) that the original v2 panel omitted. See data/sp500_pit/ for the full dataset and methodology, and experiments/monthly_dca/v5/spx_pit/IMPROVEMENTS.md for the K=3→K=2 sweep, MC delisting validation, cross-universe generalization (NDX 8/8 beats QQQ), and the rule-based rebalance sweep.

Repository layout

crt/
├── README.md                                ← this file
├── api/                                     Vercel serverless API endpoints
├── data/                                    Top-level data hub (entry-point READMEs)
│   ├── README.md                            data layer overview
│   └── sp500_pit/                           PIT S&P 500 dataset (May 2026)
│       └── README.md
├── experiments/                             experiments by family
│   └── monthly_dca/                         the production strategy family
│       ├── cache/                           cached prices, features, predictions
│       │   ├── prices_extended.parquet        daily prices (1833 tickers, the
│       │   │                                  original biased panel)
│       │   ├── features/                      per-month 79-col feature snapshots
│       │   └── v2/sp500_pit/                  PIT membership + PIT-corrected
│       │       │                              dataset and outputs (see data/sp500_pit/)
│       │       ├── sp500_membership_monthly.parquet
│       │       ├── prices_extended_pit.parquet
│       │       └── augmented/                 PIT-corrected outputs
│       ├── v2/                               GBM walk-forward training
│       ├── v4/, v6/, v7/, v8/                successive validation gauntlets
│       └── v5/                               currently deployed strategy
│           ├── score_*.py                    Chronos / time-series forecasters
│           ├── build_webapp_v5_pit.py        production v5 webapp builder
│           ├── score_winner_v5.py            today-only scorer for live picks
│           └── spx_pit/                      May 2026 PIT validation work
│               ├── REPORT.md                 full PIT validation report
│               ├── build_sp500_pit_prices.py Phase 1: augmented daily panel
│               ├── build_monthly_clean_pit.py Phase 2: monthly clean
│               ├── cache_features_pit.py     Phase 3a: base features
│               ├── add_alpha_features_pit.py Phase 3b: alpha+novel features
│               ├── build_panel_pit.py        Phase 3.5: cross-section
│               ├── build_sp500_pit_panel_aug.py Phase 5a: joined panel
│               ├── train_ml_pit.py           Phase 4: GBM retrain
│               ├── score_chronos_aug.py      Phase 5c: Chronos on augmented
│               ├── run_pit_filter_backtest.py Phase 5: k=15 baseline
│               ├── run_v3_winner_aug.py      Phase 5b: deployed v3
│               └── run_v5_winner_aug.py      Phase 5d: deployed v5
├── reports/                                 polished reports
│   ├── executive_summary.md
│   └── final_validation.md                  exp_02 winner (PIT update May 2026)
├── research/                                research notes / audits
│   └── 01_engine_audit.md                   leakage + survivorship audit
├── strategy/                                shared strategy / feature library
│   └── features/novel_features.py           12 novel features (cst, rbi, …)
├── strategies/                              other strategy families
├── crypto/, max/, spreads/                  separate strategy families
└── docs/                                    long-form research docs

How to reproduce the PIT validation

# 1) Augmented daily panel (~5 min; downloads FNSPID ~590 MB)
python3 experiments/monthly_dca/v5/spx_pit/build_sp500_pit_prices.py

# 2) Monthly clean panels
python3 experiments/monthly_dca/v5/spx_pit/build_monthly_clean_pit.py

# 3a) Base features (~65 min on CPU)
python3 experiments/monthly_dca/v5/spx_pit/cache_features_pit.py

# 3b) Alpha + novel features (~30 min on CPU, idempotent)
python3 experiments/monthly_dca/v5/spx_pit/add_alpha_features_pit.py

# 3.5) Cross-section dataset
python3 experiments/monthly_dca/v5/spx_pit/build_panel_pit.py
python3 experiments/monthly_dca/v5/spx_pit/build_sp500_pit_panel_aug.py

# 4) Walk-forward GBM (~9 min on CPU)
python3 experiments/monthly_dca/v5/spx_pit/train_ml_pit.py

# 5a) Chronos-bolt-tiny on augmented panel (~1.5 min)
pip install torch chronos-forecasting
python3 experiments/monthly_dca/v5/spx_pit/score_chronos_aug.py

# 5b) Deployed v3-winner with PIT correction
python3 experiments/monthly_dca/v5/spx_pit/run_v3_winner_aug.py

# 5c) Deployed v5-winner (with Chronos) with PIT correction
python3 experiments/monthly_dca/v5/spx_pit/run_v5_winner_aug.py

Total wall-clock: ~2 hours from a clean checkout, mostly Phase 3a (the base-feature compute).

Deployment

The web frontend is at dailystockguide.com. Deployed via Vercel using vercel.json and the serverless functions in api/. Daily refresh is handled by experiments/monthly_dca/v5/cron_daily_refresh_v5.py.

License

Internal research repository. Third-party data:

• yfinance: Yahoo Finance's terms of service apply.
• FNSPID dataset: CC BY-NC 4.0 (research / non-commercial use).
• fja05680/sp500: MIT.
• Chronos-bolt-tiny: Apache 2.0.