Running an automated book is brutally hard. The robustness needed for backtesting and automation alone is a full-time job. A month of iterating with AI loops plus my own input produced two small but real multi-year edges, and on their own they aren't worth much. AI will hand you insane-looking strategies if you judge by the backtest, and they are almost guaranteed to be overfit. If a month of AI iteration hands you a Sharpe 2 automated portfolio, it's lying. Split your data into two buckets: five years the AI can research on, and eight months kept sealed. You'll see the same pattern most of the time: beautiful numbers on the five accessible years, invalidated on the eight sealed months. Finding true edges worthy of automation was way harder than I thought, and required a lot of human input and trader experience to guide my AI so that it didn't mine good-looking numbers without logic. That experiment taught me five things: 1) There is still edge in human discretion. 2) Simple edges erode, FAST. 3) Combining discretionary angles with statistics, then eliminating (what NOT to do), might be my biggest strength going forward. 4) Anything easily mineable, like simple candlestick and volume patterns, is dry if traded alone. 5) The harder something is to turn into a testable hypothesis, the more edge it can hold. Your personal read on news, what impact a headline should have versus what it actually does, and what that implies in different scenarios, is one example of discretionary experience. Or your read on who is currently trading the market, who holds the bag, which counterparty is active, and what unwritten things bother you about a market or feel overblown in the moment. I believe there's alpha in even simple statistics, which is why I can't post it all. The method though is free: treat a stat like August seasonality as one confluence, not a signal. Alone it's trivia. Stacked with your setups it's edge. Put simply: the more data points you combine from personal experience, elimination processes, and statistics, the more often you'll find the occasional profitable trade, the kind machines are unlikely to steal. To end this on a generous note, here's a little insight into August seasonality: