US-equity quantitative research, backtesting, and paper-trading system. Codename Plutus.
Sibling of the A-share system hermes-quant: the same staged-pipeline philosophy and reusable research core, with a US-specific data / friction / execution layer. Kept as a separate repository (not a hermes subpackage) because the friction model, data adapters, universe, and execution are all market-specific.
A rigorous, survivorship-free research platform with one strategy on a forward out-of-sample watch and no live capital deployed. The market-agnostic research core (factor evaluation, walk-forward ML combiner, cross-sectional backtest engine, position sizing, idempotent paper ledger) and the US data / friction layer are implemented and unit-tested (143 tests).
Headline of the research program: after every classic family was tested under survivorship-free, cost-aware, look-ahead-audited rigor, only one edge survived as a deployable core strategy — a net-payout / buyback tilt in liquid mid/small-cap stocks — and it is now being forward paper-traded out-of-sample, not yet validated. One further edge is retail-operable but too small and episodic to be a core: the S&P 500 index delete-reversal, kept as a satellite.
The negative results are kept on the record, not buried. The major families are below; the remaining studies are written up in docs/.
-
Survivorship bias, quantified (docs/survivorship_study.md) — on a survivorship-free CRSP lake (total-return, delisting-aware prices + point-in-time membership), the candidate strategy's CAGR falls from a biased ~21% to a true ~8%, ~14 points of drawdown reappear, and the full history exposes a −88% 2008 near-ruin the biased data hid.
-
Multi-factor + regime overlay (docs/multifactor_study.md) — the regime filter is crash insurance (−88% to −43% in 2008) but whipsaws in bull markets; a passive S&P 500 buy-and-hold beats every long-only variant on CAGR, Sharpe, and Calmar.
-
Market-neutral long-short (docs/longshort_study.md) — every classic factor's net-of-cost spread is approximately zero or negative: no tradeable edge in classic factors on large-cap US (the known "factors are arbitraged" result).
-
Small-cap (docs/smallcap_study.md) — on a broad 11,219-stock survivorship-free universe, every price factor is negative at realistic small-cap costs; the only gross-positive one (reversal) is a turnover illusion.
-
PEAD, including IBES analyst surprise (docs/ibes_pead_study.md) — the drift is real (clean, monotone CAAR; IBES gives the largest +0.93% / 60d) but not tradeable net of costs. An early Sharpe-7.6 result was caught and retracted as a look-ahead artifact (announcement-gap capture); the fix (skip the reaction day) de-leaked all earlier event-time PEAD numbers.
-
ML / temporal-DL cross-sectional zoo (docs/ml_zoo_study.md) — a GRU small-cap market-neutral signal looked tradeable in-sample (~Sharpe 0.6–0.67), but its information coefficient, real in 2010–2019 (t=3.4), decayed to statistical zero by 2020–2024 (t=0.57) before the 2025 holdout printed the weakest reading on record. Verdict: on-watch, do not deploy. The most detailed negative result in the program.
-
S&P 500 index reconstitution (docs/index_effect_study.md) — the ADD run-up is dead post-effective, but the DELETE-reversal is real: dropped names earn ~+3.4% to +5.1% abnormal return net of a ~0.30% round-trip over 20–60 days. Retail-operable, but only ~11 deletions a year in distressed, volatile names: a genuine satellite, not a core strategy.
-
Net-payout / buyback (docs/issuance_study.md, docs/paper_trading.md) — the one deployable core. A long-only top-50 book in the liquid mid/small-cap band clears the buy-and-hold bar in-sample (Sharpe ~1.14, 2005–2025), signal-specific and cost-robust. It is frozen as the deployed spec (
live/strategy.py) and paper-traded forward from 2026-01-02; the first ~6-month out-of-sample read lags the small-cap index, so it is on watch, not validated. -
Pre-registered outside claims (docs/topgainer_study.md, docs/sp1_study.md, docs/biotech_catalyst_study.md, docs/copycat_13f_study.md) — four claims from a friend, each frozen in a public issue (#1–#4) before any code existed. Daily top-gainer rotation is rejected at every reading (gross −29%, net −99.5% over 2005–2024 vs +724% for the benchmark). "Always hold the largest market cap" beats on raw return (+1210%) and DCA but fails risk-adjusted (Sharpe 0.60 vs 0.65), with all of its outperformance confined to the 2015–24 mega-cap regime — rejected under the frozen rule. Biotech catalysts: across 1,257 overnight gaps ≥ +20% (2005–2024, survivorship-free), buying after the news earns −4.30% abnormal over 20 days net (clustering-robust t = −3.90, hit rate 40%) and −7.9% over 60 days — the move is over before you can act, and then it bleeds. The money moves before the announcement (+9.8% run-up), which is visible only in hindsight. An earlier run of this study was wrong and is retracted in the write-up. The drift is not about drugs (docs/gap_lottery_study.md): non-biotech stocks that gap ≥ +20% bleed the same −3.0% (t = −4.54), and the matched biotech-minus-control difference is +0.28% (t = 0.01) — so the biotech "sell the news" claim is rescoped into a broader and sturdier one, do not chase any +20% gap, worse for small caps and for bigger jumps. Copying the greats' 13F filings: 17 hindsight-selected legends (a God's-eye view the hypothesis could never have had), entered at the filing-date close — no edge at any horizon, and once size-matched to the benchmark the abnormal return is +1.1%/yr with t = 0.6, indistinguishable from zero. Patience does not rescue it; Buffett's own new positions did not beat the market either.
Also written up in docs/: the first point-in-time factor read; event-time PEAD, small-cap PEAD, short-term reversal and overnight returns — each a real gross effect that is not retail-tradeable net of cost; pairs trading, whose mean-reversion alpha is no longer there at all; and volatility-managed exposure, a genuine risk-adjusted improvement that is an overlay, not alpha.
The durable asset is the methodology — survivorship-free, cost-aware, and look-ahead-audited — that reports the truth regardless of the result.
Offline research and online execution are separated deliberately. A strategy advances a stage only when the prior stage holds up:
+-----------------------------+ +--------------------------+
| RESEARCH (offline) | signals | EXECUTION (online) |
| local PC (Windows) | ------> | local PC (Windows) |
| | (files) | |
| factors / ML combiner | | EOD paper ledger |
| friction-faithful backtest | | -> (later) Alpaca live |
+-----------------------------+ +--------------------------+
|
v
(optional) zipline-reloaded friction cross-check
($0 commission / SEC Section 31 fee / FINRA TAF / slippage)
- Backtest on historical data, offline, through a US-faithful friction model
(
research/backtest/frictions.py). An optional independent cross-check viazipline-reloadedplays the role RQAlpha plays for hermes. - Paper trading (simulated): a lightweight idempotent end-of-day ledger (
live/ledger.py) that replays the SAME research engine forward, so there is no train/serve skew. A monthly-rebalance strategy needs only an EOD feed. - Live (small real capital): deferred. Same strategy object, with an Alpaca gateway swapped in.
US vs A-share — what changes, what does not. The research core (eval / model / factors / backtest mechanics / sizing / IO) is market-agnostic and shared with hermes by copy-and-diverge. The market-specific layer is rebuilt: data vendors, the friction model, the universe, and execution. Crucially, US frictions are much lighter — $0 commission, no stamp tax, no 100-share lot, no daily price limit, no T+1 holding lock — so the dominant risk shifts from frictions (the A-share crown jewel) to data quality: survivorship bias and a point-in-time universe (see docs/data_sources.md).
The exact US market rules and fee rates baked into the friction model are tracked, with primary-source citations and as-of dates, in docs/MARKET_FACTS.md.
A dedicated conda environment plutus (Python 3.12), separate from the hermes environment
because the dependency stack differs (yfinance / SEC EDGAR / Alpaca, not BaoStock / Tushare).
conda create -n plutus python=3.12 -y
conda activate plutus
pip install -e ".[dev]" # research core + the pytest suite
cp .env.template .env.local # then fill in any keys you use (all optional to start)
pytest # the research-core unit suite
python scripts/probes/smoke_yfinance.py # verify the free data link
Secrets are read from the real environment first, then from .env.local (gitignored; never
committed). yfinance and Stooq need no key, so research runs with nothing configured. Optional
extras: pip install -e ".[broker]" (Alpaca) and pip install -e ".[notebooks]" (JupyterLab).
Free-tier first (the US analog of hermes starting on anonymous BaoStock):
| Source | Auth | Role |
|---|---|---|
| yfinance (Yahoo) | none | free daily adjusted OHLCV backbone — the default |
| SEC EDGAR | free User-Agent header | official fundamentals for value / quality factors |
| fja05680/sp500 | none | point-in-time S&P 500 membership (free universe) |
| CRSP (via WRDS) | bring-your-own licensed extract | survivorship-free total-return prices + PIT membership; not included |
| Stooq | none | daily CSV cross-check; endpoint currently behind a JS bot-check (best-effort) |
| Alpaca | free key | paper-trading account + market-data API (registered, parked) |
Survivorship bias is the main free-data weakness: Yahoo drops delisted tickers, and a free point-in-time membership history is hard to assemble cleanly. The survivorship-free backtests use a licensed CRSP extract that is not part of this repository (see Disclaimer); see docs/data_sources.md for the approach and its limits.
src/plutus/ the engine — importable package (src-layout); no trading-framework dependency
config.py secret / token loading (environment -> .env.local)
paths.py, io.py on-disk locations; atomic file writes
data/ sources/ (yfinance prices, SEC EDGAR fundamentals, CRSP, IBES, Stooq);
universe.py (point-in-time S&P 500 membership) -> adjusted parquet lake
research/
backtest/ friction-faithful backtest: frictions (US), portfolio, long-short, pairs,
optimize, regime, sizing, shared metrics
factors/ factor library (value, reversal, momentum, low-vol, quality, net-payout)
eval/, model/ single-factor IC + calibration; walk-forward LightGBM combiner
live/ EOD paper trading: deployed net-payout spec (strategy.py), idempotent ledger,
CRSP replay (paper.py), free-data forward read (forward.py), data feed
execution/ Alpaca live-gateway adapters — deferred stub, unused
scripts/ research studies (*_study.py), data-lake builders, and paper-trading drivers; probes/
tests/ pytest suite: engine invariants, no-look-ahead, friction model, ledger parity
data/ local data lake — INPUTS (gitignored)
results/ generated OUTPUTS: signals, backtests, figures, paper ledgers (gitignored)
external/ upstream checkouts (zipline-reloaded), pip install -e — gitignored, unmodified
docs/ architecture, market facts (cited), and curated research findings (tracked)
notebooks/ research scratch (gitignored)
This is a personal research project. Nothing here is investment advice. The strategies are research hypotheses under test, not recommendations, and no live capital is deployed. The CRSP and IBES data used for the backtests is licensed for personal research use only; it is not included in this repository and must not be redistributed — supply your own licensed extract. The code is provided as-is, without warranty of any kind.