Skip to content

paper_trading.md

Latest commit

 

History

History
144 lines (114 loc) · 9.28 KB

paper_trading.md

File metadata and controls

144 lines (114 loc) · 9.28 KB
Raw

Forward paper-trading the deployed net-payout strategy (the out-of-sample gate)

docs/issuance_study.md found the one retail-operable edge plutus has: net-payout / buyback in liquid mid/small-cap US stocks (long-only top-50, ADV > $5M/day), Sharpe ~1.14 full / ~1.40 since 2021 net of 15bps, vs a cap-weighted buy-and-hold bar of 0.61. That result is in-sample (2005-2025) and its magnitude is above the published issuance literature, so it has not earned real capital. This module is the gate it must clear first: a forward, out-of-sample paper record, the same discipline the sibling hermes-quant A-share book is held to.

Architecture — the research engine IS the strategy brain (no train/serve skew)

Train/serve skew (the served strategy quietly differing from the researched one) is the dominant silent alpha-killer. It is eliminated here by construction:

  • One signal definition. plutus.research.factors.library.net_payout is the single source of truth. Both scripts/crsp_issuance_study.py (research) and live.strategy.deployed_signal (serving) call it. tests/test_paper.py::test_deployed_signal_is_the_net_payout_factor locks this; the dry-run proves the deployed spec reproduces the study headline to the digit.
  • One universe definition. live.strategy.deployed_members is the cap-rank band [500, 3000) intersected with the ADV > $5M/day liquidity screen — the exact universe of the study's headline.
  • One decision engine. live.paper.replay runs the research backtest (signal_portfolio_backtest(..., collect_trades=True)) and folds its per-fill trade log, day by day, into an idempotent live.ledger.LedgerState valued by the SAME valuation_panel. The ledger only records the engine's decisions; it never re-decides.

The anti-skew parity gate (scripts/paper_dryrun.py) asserts the ledger equity equals the engine equity bar-for-bar, and that the deployed spec reproduces the documented headline:

seed $1,000,000: parity max|ledger-engine| = 5.96e-08 (rel 6.5e-16) -> OK
headline @ $1M (vs docs/issuance_study.md): CAGR +23.7% / maxDD -32.5% / Sharpe 1.14 -> OK

The deployed spec (frozen as of this commit; live.strategy.DEPLOYED)

field value rationale (docs/issuance_study.md)
signal net-payout, 1y (252d) the post-publication-robust, low-turnover, survivorship-free anomaly
book long-only top-50, equal weight the realizable retail form; top-30/50/100 all Sharpe 1.33-1.52
universe cap-rank band [500, 3000) ∩ ADV > $5M/day the liquidity-screened (tradeable) headline universe
rebalance monthly, no turnover buffer low-turnover signal does not need a band
cost $0 commission + 15bps slippage the validated headline (50bps stress still clears B&H at 1.07)

Inception and the out-of-sample discipline

PAPER_INCEPTION = 2026-01-02. The study used the full 2005-2025 sample (the "2025 holdout" was reported, so 2025 is in-sample); the first genuinely out-of-sample bar is the first trading day of 2026. The seed is invested at the first available close ≥ inception into the then-current top-50, and total_return / max_drawdown are measured from there. The spec above is frozen as of this commit and must not be refit on post-2025 data — refitting on the holdout would destroy the gate.

Data constraint (the key difference from hermes-quant)

hermes auto-refreshes from free BaoStock. plutus runs on CRSP — a paid, manual pull with no free daily feed, so this module does NOT auto-refresh; it replays whatever is on disk. The lake currently ends 2025-12-31, before the inception, so every tier honestly reports status="awaiting_data" (seeded, no positions) rather than silently falling back to the in-sample backtest:

tier      label  status         as_of       n_bars  equity     totRet  ...
25,000    small  awaiting_data  2025-12-31   0       25,000     +0.0%
...
seeded; lake ends 2025-12-31 < inception 2026-01-02 -- land a fresh CRSP pull
covering post-inception bars, then re-run to extend the record.

To advance the forward record: land a fresh CRSP pull that includes 2026 bars (the build scripts), then re-run scripts/paper_live.py. The ledger is idempotent (recompute-from-seed), so re-running any date reproduces it.

Running it

conda activate plutus
python scripts/paper_dryrun.py                     # parity gate + headline reproduction (CI)
python scripts/paper_live.py                       # paper mode, all tiers, persists to results/paper/
python scripts/paper_live.py --tiers 100000 1000000
python scripts/paper_live.py --as-of 2026-03-31    # cut the lake at a date (idempotent re-run)
python scripts/paper_live.py --backtest            # archive the full-history backtest to results/backtests/

Each run persists, per tier, paper_curve_<seed>.parquet (equity), paper_trades_<seed>.parquet (full fill log), and paper_report_<seed>.json (the snapshot below) under results/paper/.

Reading the forward record

The report benchmarks the strategy against the cap-weighted buy-and-hold over the same forward window (bh_total_return, bh_ann_sharpe) — the out-of-sample version of the study's comparison. ann_sharpe is null until ≥ 21 forward bars accrue (a Sharpe on a handful of days is noise). Judge the strategy by whether it clears the B&H bar OOS, not by its absolute return.

A machinery smoke test on the in-sample 2025 window (NOT the forward record) confirms the active path reproduces the studied edge on real data: +45.4% vs B&H +18.3%, Sharpe 1.81 vs 0.99 — i.e. the live code yields the same strong-recent-years result the study found.

Free-data early read — start now, before the next CRSP pull (live.forward)

Because the CRSP path is awaiting_data, live.forward gives an immediate, lower-rigor forward read from FREE data, without waiting for a paid pull. The split keeps it honest: the book is selected from clean CRSP at the last lake bar (survivorship-clean decision, validated signal + universe), then priced forward with yfinance adjusted closes and benchmarked against small-cap ETFs (VB, which tracks the CRSP US Small Cap index; IWM, Russell 2000). It tests the selection (do the picked names beat the index OOS), not the monthly rebalance — net-payout is low-turnover, so a held book approximates the strategy between CRSP refreshes.

Coverage is complete: of the 50-name 2025-12-31 book, 47 have direct free 2026 data and the other 3 are handled by a curated, primary-sourced corporate-action resolver (CORPORATE_ACTIONS in live.forward) — none is silently dropped (dropping would renormalize weight onto the survivors, a survivorship-y bias):

name 2026 corporate action (primary-sourced) forward treatment
CommScope COMM sold CCS unit to Amphenol; renamed Vistance Networks, → VISN; $10/sh special dividend priced as VISN (incl. the dividend)
Denny's DENN taken private by a TriArtisan consortium, $6.25 cash/share realized at the cash deal → flat
First Foundation FFWM all-stock merger into FirstSun, → FSUN (0.16083) priced as FSUN

Run: python scripts/paper_forward.py.

First read (2026-01-02 .. 2026-06-23, 118 bars, 50/50 priced, seed $1M) — a caution flag, not a verdict:

book total return maxDD Sharpe
net-payout top-50 (held) +5.0% −15.8% 0.60
benchmark VB (CRSP small-cap) +13.2% 1.60
benchmark IWM (Russell 2000) +19.2% 1.93

The book lagged the small-cap index by 8-14 points over the first ~6 OOS months — the opposite of the in-sample story (where it beat the B&H bar). Read this with heavy caveats: (1) the sample is thin (~6 months — the Sharpe is noise-dominated, read direction not magnitude); (2) it is selection-only (no forward rebalance); (3) early 2026 was a sharp small-cap beta rally that a value- tilted net-payout book would be expected to trail; (4) VISN's +40% reflects the $10 special dividend on the standard reinvested-total-return basis (consistent with the study's CRSP convention), which slightly flatters one 2%-weight name. It is not the definitive test, but it is the first real out-of-sample evidence, and it does not corroborate the in-sample edge. The monthly-rebalanced CRSP ledger (paper_live.py), run when a fresh CRSP pull lands, remains the decisive gate.

Honest caveats (carried from docs/issuance_study.md, plus the feed constraint)

  • Magnitude is above the literature (~0.5-0.8 published issuance Sharpe). The forward record is expected to come in below the in-sample 1.1-1.4; treat the in-sample level as an optimistic upper bound and watch whether it merely clears B&H, not whether it matches 1.4.
  • One searched signal. The IC t=4.4 and cross-signal controls make this far more than a lucky fit, but the forward test is precisely what converts "survived the in-sample controls" into evidence. Do not size real capital until the forward record holds up.
  • Long-only ≈ beta + a tilt, carrying full small-cap market beta; pair with the capped volatility overlay (docs/vol_overlay_study.md) if the drawdown is to be managed.
  • No live feed. Because CRSP is a manual pull, the forward record advances only as fast as new pulls land — the lake_lag_days field flags how stale the lake is on each run.