Thirty-One Signals, Six Bets · overview and contents
Null result · Quantitative trading
A retail account cannot assemble enough independent bets to detect an edge it might actually have — and adding more price signals does not help.
A Wall Street Journal profile described an investor delegating capital to named LLM agents. We built a purpose-designed backtesting and paper-trading harness to ask whether that idea, or any of its rules-based neighbours, survives honest measurement in a retail-sized account. Nine experiments were run over a decade of daily bars: two LLM agent designs, adaptive strategy selection, transaction-cost measurement, a small-cap capacity test, a turnover sweep, a pre-registered out-of-sample test, a market-neutral construction, and a breadth study.
No configuration beat buy-and-hold. The central result is structural rather than circumstantial: 31 signals drawn from six families of the published anomaly literature collapse to 5.6–6.2 independent bets. Zero signals reach |t| ≥ 2 in large caps and one does in small caps, against 1.6 expected by chance. The combined information coefficient is negative in both universes. At the measured IC, a usable information ratio requires roughly 1,300 independent bets per year; 31 price-derived signals supply about 67.
The shortfall is two orders of magnitude, so it is not closed by writing more signals from the same price series. We also argue that the one remaining path — LLM-extracted features from text — cannot be backtested honestly, because the contamination lives in the model's weights rather than in the data handed to it.