Research
Research that is allowed to say no.
Our research process is built to find out when an idea does not work. Below is the workflow, where the data comes from, and an unvarnished summary of what our own tests found.
Workflow
Five steps, every time
01
Ask
A hypothesis is written down before data is examined: universe, signal timing, horizon, costs, benchmark.
02
Source
Inputs come only from data with a recorded publication time; the research council cites sources and as-of times.
03
Test
Purged walk-forward evaluation with a locked final holdout; every trial — including failures — is registered.
04
Doubt
Bootstrap confidence intervals, deflated Sharpe and multiple-testing correction decide whether an effect is distinguishable from luck.
05
Decide
Only a named human can promote a strategy; evidence and caveats are recorded with the decision.
Findings
What our tests found
No model significantly beats equal weight
Factor rank, 12–1 momentum, ridge regression and gradient-boosted ranking were tested out of sample (2020–2025). None beat holding the universe equally weighted with statistical confidence; machine learning added no value over simple baselines.
Strategy pods: edges not significant
Re-examining stored pod validations with confidence intervals, no pod's Sharpe ratio is significantly above the equal-weight universe. Several (insider buying, quality, low volatility, mean reversion) are significantly worse.
Headline tone has no predictive power
Across 334 weekly decisions, the cross-sectional rank correlation between 3-day news tone and forward returns was indistinguishable from zero.
Risk overlays trade return for drawdown
A trend/volatility/breadth exposure overlay reduced maximum drawdown materially, but its risk-adjusted improvement was not significant while its return cost was.
Provenance & limitations
Provenance
Every research input records its provider and the time it became public. Council theses store a data-snapshot hash and per-agent outputs; the audit log links each to the proposals it produced.
Limitations
- The stock universe is today's ~100 large caps, which flatters history (survivorship bias). Licensed delisted-security data would be required to remove it.
- No licensed historical analyst-consensus data, so earnings-revision signals are not studied.
- Daily bars only; no intraday or order-book history.
- Language-model analysis can be wrong or inconsistent; it never sets position sizes.
Methodology details: backtesting · strategy-by-strategy evidence: strategies.