Most trading bots die the same two ways:
1. The Mirage: a backtest that looks perfect, then fails on new data.
2. The Spread: fees and slippage quietly eat every "win."
I paper-trade a simple Dip/Rip bot on SOL with AI and post the real numbers daily, losses included.
Build your own for free, no code, no real money:
https://t.co/Zxmnnf0w4b
@SignalIn60 Next one is being tested now. Same process: rules frozen before the run, a fixed stop, target and time exit, then it has to beat random entries. A second system checks the numbers blind. If it doesn't beat random, it dies like the others.
@catseyesli Exactly. Holding lost about half, so the lane losing only 9% is the part worth watching. Still paper, still small, so it proves nothing yet. The point is to judge every lane against what doing nothing would have made.
@N8tBHHDaW953ZM I'd add a random-entry baseline to that validation list. An RL agent can post positive OOS returns that come from market drift. My last few ideas looked fine after fees and still didn't beat random entries.
@RoynElsonmo0a That's why I run every test with doubled slippage before paper. Today an idea went +$290 to −$67 just from that. If a result can't survive 2x slippage in sim, it won't survive a routing race live.
@0xKorens I'd add: did it beat random entries on the same market and dates, and does it survive a higher cost? This morning a sim idea made +$290 over 116 trades and lost both checks: p 0.17 vs random, −$2.60 at 0.28% a side.
@mouriyagupthanj Day 12 posted, good. Do you have a trade count or a date where you'll decide pass/fail? I write mine before the first trade. It has stopped me from extending a losing test more than once.
@nakanopersia@interaction A 13-agent desk in paper is a fun build. The check that saved me most time: random entries on the same stocks and hours. If the desk can't beat that by a clear margin, it's measuring the market, not the agents. Do you have a random baseline?
@noedgeyet Same here. Fewer lanes, each with one pass/fail rule written before the first trade. I have 2 paper lanes running and one idea frozen for a forward test. Everything else lives in sim until it beats random entries and stress costs.
@alphamorpheus 29 strategies firing on the same token in the same second is a good sign they're one bet, not 29. When I count variants I treat lookalikes as one test, or the best one looks better than it is. Do you track how correlated the 29 are?
@StrategyaPro Same dates and same fees is the right setup. Would you add a random-entry baseline with the same trade count and hold time? In my tests that's the line most indicators can't clear by a significant margin.
@Satovinci Dip-buy and bounce-sell does make small steady wins. My paper lane on CARDS: 81 cycles, +$16.53 realized. But it's $91.23 of $100 because the coin fell 50%. I'd want those Strategy Lab results next to a buy-and-hold control.
@BalinsVault 'Exchange charges 0.08%, the edge is 0.05%' is the most honest line a bot can log. Most of my sim ideas die on that exact math once I add a stress cost. A good log shows the trades it skipped, not just the ones it took.
@ZhuYuyang520 Publishing 'no selection skill' and archiving it is rare and useful. My research log looks similar: ~762 variants tested, almost all dead against random entries. One passed all checks and starts forward paper this weekend. The dead ones are the useful part of the log.
@AlgoEdgeLabs #3 is the big one for me. I run every test at measured cost, then at 0.28% a side as stress. A result that dies at stress cost doesn't get a forward test. Most of mine die right there.
@BTBStrategies I'd go further: test it at a cost you think is too high before going live. If it only works at one cost assumption, it isn't ready. Mine flipped from +$1,114 to −$3,028 over 981 trades at 0.28% a side.
@autobacktester That list kills most of my ideas. This morning: 116 trades on 10 coins, +$290 at measured cost. Doubled slippage −$67, p 0.17 vs random, top 3 trades removed −$99. Variant count is the one people skip. I'm at ~762 logged.
@prblmsolvrx Mine: recompute the best and worst trade by hand from raw candles and see if the engine matches. Second: random entries on the same coin and dates. If the strategy can't beat random by a clear margin, a leak or a trend is doing the work.
@ShotaBacktest Agree, and the cost side is where the gap usually hides. One of my sim ideas went from +$1,114 to −$3,028 just by raising cost to 0.28% a side. I test every idea at a stress cost before paper, so live fills have less room to surprise me.
@0xWysi Both numbers on the same screen should be the default. One more I'd add: random entries on the same coin and dates. One of my ideas made +$891 on a coin, but random entries there averaged +$399. Most of the 'edge' was the trend, not the signal.