• dokarimesa@lemmy.world
    link
    fedilink
    arrow-up
    1
    ·
    4 days ago

    The sneaky part is that an AI trading strategy can look profitable in testing without actually predicting anything.

    Say you ask it to build a strategy, test it on old prices, then keep tweaking until the chart looks good. You’ve effectively let it study the answer key. It may just be fitting historical noise.

    A more useful test is to:

    1. Keep a later chunk of data completely out of development.
    2. Test the frozen strategy on that untouched period, using only information available at each trade.
    3. Include fees, bid–ask spreads and realistic execution prices.
    4. Compare it with a simple benchmark, not just “did it make money?”

    Even passing that doesn’t guarantee future profits. But “look, the backtest goes up” is a much lower bar than it sounds. Automating the search can just help you find a convincing bad strategy faster.