Written by an AI agent working in this repository. Hypothetical backtests, not investment advice.
The request was "automated agentic trading, maybe through Robinhood, like a fund, with lots of backtesting and advanced strategies." The hard part of that is not getting an agent to place orders. It is stopping an agent that can try a thousand ideas from finding the one that looks great by luck, and then trading it with real money.
So Agentic Trading Lab treats every idea the agent tries as a cost:
- It proposes configurations of eight published rules (trend, time-series and cross-sectional momentum, dual momentum, inverse volatility, reversal) from a bounded search space.
- It backtests each on data up to 1999 only, with costs, and logs it whether it looks good or not.
- It judges the best one with the Deflated Sharpe Ratio (Bailey and López de Prado, 2014). This raises the bar with every trial. It also uses the Probability of Backtest Overfitting (Bailey, Borwein, López de Prado and Zhu, 2017), which asks how often an in-sample winner lands in the bottom half on unseen slices of the same data.
- It runs the winner once on 2000 to September 2026, then gates it: DSR ≥ 0.95, PBO ≤ 0.5, holdout Sharpe positive with 90% confidence.
What happened
| Universe | Winner | Holdout Sharpe (vs equal weight) | Drawdown (vs EW) | DSR | PBO |
|---|---|---|---|---|---|
| 10 US industries, 1926 on | dual momentum 12m, top 2 | 0.50 (0.53) | −27.9% (−48.1%) | 0.999 | 0.50 |
| 49 US industries, 1969 on | industry momentum 12m, top 5, vol-targeted | 0.62 (0.55) | −21.8% (−52.7%) | 0.890 | 0.19 |
| BTC + ETH daily, 2016 on | 20-day moving average | −0.08 (0.31) | −50.5% (−68.8%) | 1.000 | 0.81 |
None passed. The 49-industry winner came closest: alpha of 3.5% a year against equal weight, Newey-West t = 2.09. But it was the best of 98 tries, and a deflated Sharpe of 0.89 means luck is not ruled out.
The crypto row is the textbook case. Its in-sample Sharpe was 2.07. The Deflated Sharpe Ratio passed it at 1.000, because DSR mostly charges for how many things you tried, and a raw Sharpe of 2 clears that bar. PBO flagged it at 0.81: in most re-splits of the 2016-2021 data, whichever rule won one half did badly in the other. After 2021, at 50 bps a trade, it lost money. A gate using only DSR would have promoted it.
What did hold up
- Trend rules cut drawdowns roughly in half. The 10-month moving-average rule (Faber, 2007) lost at most 51% over 1929-2026 on 10 industries against 82% for equal weight, and 27% against 53% on 49 industries. Its Sharpe was higher too, but most of the value is in the drawdown, not the return.
- Slow beats fast after costs. On crypto the 200-day rule kept a holdout Sharpe of 0.70 at 50 bps while the 20-day rule went negative. One-month industry reversal fell from 0.40 to −0.05 between 0 and 50 bps.
- Re-picking the best rule every year by trailing Sharpe matched equal weight (0.62 vs 0.62 since 1939 on 10 industries), with a much smaller worst loss.
About "like a fund"
The software is for one person trading their own account, and it says so in its help text, page and LEGAL.md. Managing or pooling other people's money brings in adviser registration, investment-company and securities-offering rules in the US. That is a lawyer's question first. Robinhood publishes an official API for crypto only, which the tool supports. Stocks and ETFs go through Alpaca, which has free paper trading. Live trading needs a flag, an exact confirmation phrase, no kill-switch file, a strategy that passed the gate, and per-order limits.
Everything here is reproducible with tradelab study. The data comes from the Kenneth R. French Data Library and Coinbase's public candles, downloaded at run time. Every number above is in results/study.json, and a test fails if the product page quotes one that differs.