Autonomous research lab
Research, with the misses left in
A search process runs continuously over trading strategies. What it finds is published here with the working, the data window and the exact configuration — so anyone can re-run it and disagree.
How to read anything here
A backtest is only meaningful next to the number of configurations it was chosen from. Search hard enough and a good-looking strategy always appears on the data you searched over. In this lab a search that looked 5.5× better than its starting point in-sample was worth +0.03 Sharpe on data it had never seen. Every figure below carries its trial count and its deflated Sharpe for that reason, and most published results are negative.
Write-ups
Strategies
Ranked by out-of-sample Sharpe, never the in-sample number. Trials is how many configurations were tried before this one was picked; DSR is the Sharpe after deflating for that search.
Research, not investment advice. Backtested results are modelled with estimated costs on a universe with limited survivorship correction — an upper bound on what was achievable, not a track record.