Jev trading bot's 1,260 decisions all lost to buy-and-hold
A Jev bot on OpenRouter made buy/sell/hold calls on BTC, ETH, SPY, QQQ and gold over 84 days; its best strategy returned +5.9% versus +12.4% for equal-weight buy-and-hold.
MY JEV BOT MADE 1,260 DECISIONS FOR $0.11 AND EVERY SINGLE ONE OF THEM WAS WORSE THAN DOING NOTHING. The setup was clean. Three months of daily investment reports, one per day, fed to Jev on OpenRouter together with the current book. For every asset it returned buy, sell or hold with a probability on each option. A separate program sized positions off those probabilities and executed on real Binance hourly candles. 84 days. Three strategies: momentum, contrarian, macro. Five assets: BTC, ETH, SPY, QQQ, gold. 1.2 seconds per call. Zero failed calls. Momentum finished at +5.9%. Contrarian at -1.4%. Macro at -9.1%. Equal-weight buy and hold over the same 84 days: +12.4%. The best strategy lost to a portfolio that never placed a trade by more than six points. The other two lost to it by fourteen and twenty one. The dashboard is the only part that went perfectly. A second model, Ling-3.0-flash-VL, got a rough sketch and built the page on its own: net value curve, position bars, decision flow. Playwright turned 420 frames into a 14-second clip of the whole quarter. So the honest read. Judgment at this price and speed is real. Edge is not. What changed is the cost of finding that out. A quarter of backtest for eleven cents means you can be wrong a hundred times before lunch and still have a budget.