Portfolio Optimizer
A systematic equity research project, run as an adversarial process: every idea is
pre-registered, tested against a declared null, and killed when it fails.
312 recorded research steps. Nineteen signal families closed. Zero strategies promoted.
Zero dollars traded.
That last line is the result, not an apology for one. A research process that promotes
nothing from 312 attempts is either broken or working exactly as intended, and the record
below is the evidence for which.
What this project set out to do
Find a systematic equity strategy with a real edge — using only free data, on a laptop,
held to the standard a quant firm would apply. The dashboard shows candidate strategies with
trailing returns from 100% to 175%. None of them is an edge, and the research explains
why.
The main findings
1. The strategies never had cross-sectional skill.
Thirteen dashboard signals and fourteen closed families were measured three beta-neutral
ways — rank information coefficient, top-minus-bottom decile spread, and decile
monotonicity. None clears Bonferroni. More decisively, none orders its deciles:
monotonicity sits within ±0.17 of zero for every signal tested. A signal with genuine content
ranks its universe; these rank nothing.
(Steps 296, 298, 299, 300)
2. The headline returns were market beta and a few lucky names.
Taking the one signal that survived out-of-sample testing and neutralising its market
exposure dropped it from 16.65% CAGR to 1.59% — roughly ninety per cent of the return was
beta. The top-N books were not winning on selection.
(Step 295)
3. Nothing survives a window it was not selected on.
Every fundamental panel began in 2023, so no strategy here had ever been tested outside its
selection window. Acquiring SEC Financial Statement Data Sets back to 2012 fixed that.
Nought of six survive; three of six have the wrong sign. Growth, the flagship, lost 15.4
points a year to its own scored universe.
(Steps 287, 289, 290)
4. Breadth is not the binding constraint — skill is.
The project spent roughly fifty steps trying to raise breadth on the theory that
IR = IC × √BR made it the lever. It is a product: an IC indistinguishable from zero yields
an IR indistinguishable from zero at any breadth. Adding a genuinely uncorrelated asset
raised effective bets from 2.00 to 2.88 and bought 0.021 of Sharpe. Every liquid domain
measured collapses — 35 ETFs give 4.16 effective assets, 20 crypto assets give 3.09.
(Steps 246, 277, 279, 285, 296)
5. A one-week date convention invalidated the project's best result.
Two artifact families disagreed about whether a return stamped at date D was the week
ending or beginning at D, and nothing recorded which. Joining them by date compared week
t against week t+1, producing near-zero correlations that looked like diversification.
Corrected, the "independent" strategy correlates +0.57 to +0.77 with the others, and a
blend that appeared to beat both components beats neither. The finding was withdrawn and the
forward clock built on it was superseded before it started.
(Steps 291, 292, 294)
6. Market-neutral structure works. There was nothing to put in it.
A long-short book removes market beta by construction — realised betas of −0.009 and +0.007,
clearing an orthogonality gate that fourteen families had failed. Both signals tested in it
had no skill. The machinery is sound and unfuelled.
(Steps 293, 295)
What was tried
Roughly forty signal families across eight categories — momentum is four of them.
| category | families |
|---|
| Price / return | short-term reversal, seasonal momentum, opening-range breakout, coskewness, idiosyncratic skewness, trend consistency, sector dispersion, vol-of-vol, downside beta, residual reversal, industry-return-of-big-firms, daily OHLCV zoo, 22-signal literature screen, cross-asset crisis trend, futures trend |
| Fundamental | growth, cash conversion, balance-sheet quality, profitability, shareholder discipline, quality acceleration, earnings yield, FCF yield, sales yield, composite value, quality-at-reasonable-price, low asset growth, accounting-change family |
| Events | PEAD / SUE, 13D & 13G activist filings, Form 4 insider clusters, Form 4 routine-vs-opportunistic |
| Positioning / flow | 13F institutional linkage, FINRA short interest, FINRA daily short-sale volume |
| Text | 10-K language change, earnings-call transcripts |
| Network | supply-chain graph |
| Other markets | 35-ETF multi-asset, 20-asset crypto, futures + roll repair |
| Structure & ML | pairs / statistical arbitrage, long-short market-neutral, risk-budgeted sizing, breadth repair, meta-labeling, triple-barrier labelling, fractional differentiation, MDA/MDI importance, market-state and macro-state classifiers |
Method
The rules are in CLAUDE.md and they are enforced, not aspirational:
- Pre-registration. Configurations, gates, signs and outcome readings are frozen in
2.0/config/ before results exist. A signal declared positive that measures negative is
reported as refuted, never re-interpreted as contrarian.
- Point-in-time or it does not count. Universes are built from historical filer rosters;
a filing enters on its filing date, never its period end.
- Multiple testing compounds across the whole project, not the current batch.
- Costs are not optional — 0, 10, 50, 100bps, and borrow on any short.
- Negative results are preserved.
2.0/PROJECT_HISTORY.md is
append-only and includes every error the author made, several of which invalidated earlier
conclusions.
On errors
Ten measurement defects were found and recorded, most by noticing a number that could not
mean what it appeared to mean. Three times in a single session a result cleared a
significance bar and dissolved under inspection: an insider signal at p=0.0000 that was a
presence test on a 93.3%-zero series; a trend signal at p=0.0017 with a negative decile
spread; a text signal at p=0.0030 sitting below its own data's resolution. A process that
stopped at the p-value would have reported all three as discoveries.
What happens next
Five forward clocks began on 2026-09-11: a residual composite, an equal-weight benchmark,
a tie-agnostic companion, an earnings-surprise book and a valuation book. Each requires 52
untouched weeks. A missed weekly window cannot be backfilled, and the historical record never
advances a clock.
This is the only source of evidence the project has left that has not already been searched.
Repository layout
1.0/ the legacy ETF project this began as -- docs, reports and
source kept as the origin record; its data is superseded by 2.0
2.0/
dashboard/ Next.js research dashboard (the deployed site)
scripts/ acquisition, construction, testing and audit scripts
src/ forward-evidence ledger, return conventions, evaluation
config/ frozen manifests, pre-registrations, forward protocols
docs/ runbooks, protocols, the research queue
evidence/ saved outputs of every experiment
PROJECT_HISTORY.md the full append-only record, 312 steps
CLAUDE.md the operating rules
Running it
cd 2.0 && python -m venv .venv && ./.venv/bin/pip install -r requirements.txt
./.venv/bin/python scripts/rebuild_all_data.py --core # panels the dashboard and clocks need
cd dashboard && npm install && npm run dev
Raw acquisition caches are not stored in the repository — they are large and every byte is
re-downloadable from SEC EDGAR, FINRA and public price sources. rebuild_all_data.py
re-acquires them; --core fetches only what the dashboard and the forward clocks read.
Point-in-time vintage snapshots cannot be re-downloaded, because they record what a source
said on a given date. Those are preserved in the repository, which is why a clone is around
400MB rather than a few tens of megabytes: the alternative was a repository that cannot
reproduce its own point-in-time claims.
scripts/slim_repository.py shows what is safe to remove and what is not, with a restore
command for every path. It defaults to a dry run.
Honest limits
- The universe is US-listed issuers; the fundamental work concentrated on technology and
energy before being widened to 2,681 issuers across 69 sectors in Step 300.
- Weekly data throughout. Several ideas — pairs convergence in particular — are properly
daily, and a negative result on weekly bars is weaker than a positive one would be.
- Short-sale borrow is modelled as a flat annual charge. Real borrow is name-specific and
worst on exactly the names a short leg wants.
- No strategy here has forward evidence. As of this writing the clocks have just begun.
Licence and status
Research only. Nothing in this repository is investment advice, and every configuration
carries live_trading_enabled: false.
