Win Rate Lost to the Exit: Four MT5 LLM Bots Closed the Day at -1,445 Yen

June 16, 2026 was not a “bad entries” day as much as an “exit discipline” day. Across the four MT5 bots, the combined realized result was -1,445 yen, even though the total win rate was still 53.3%. That number made me pause a bit, because more trades won than lost, and yet the day still bled out.For this series ledger, I am treating the pasted history as Day 1: cumulative realized P/L is -1,445 yen. Open unrealized P/L was not included in that total. BoundSniper still had -30 yen floating, and MLScore GF-T4 GB had -124 yen floating, so the unfinished part of the day was not exactly clean either.Bot-by-bot results■ GateGrid AI -703 yenPair: GBPUSD-Closed trades: 4Record: 2W / 2LWin rate: 50.0%Gross profit: +93 yenGross loss: -796 yenPayoff ratio: 0.12Max loss: -432 yen■ BoundSniper +3 yenPair: USDJPY-Closed trades: 3Record: 2W / 1LWin rate: 66.7%Gross profit: +162 yenGross loss: -159 yenPayoff ratio: 0.51Max loss: -159 yen■ MLScore GF-T4 GB -487 yenPair: GBPJPY-Closed trades: 4Record: 2W / 2LWin rate: 50.0%Gross profit: +351 yenGross loss: -838 yenPayoff ratio: 0.42Max loss: -600 yen■ LLMBridgeTrader -258 yenPair: EURUSD-Closed trades: 4Record: 2W / 2LWin rate: 50.0%Gross profit: +135 yenGross loss: -393 yenPayoff ratio: 0.34Max loss: -249 yen■ Total -1,445 yenScope: realized closed trades onlyClosed trades: 15Record: 8W / 7LWin rate: 53.3%Gross profit: +741 yenGross loss: -2,186 yenPayoff ratio: 0.30Max loss: -600 yenToday’s theme: the exit beat the win rateThe headline is not just the loss. The stranger part is that the bots won 8 out of 15 closed trades and still ended at -1,445 yen. A total payoff ratio of 0.30 means the average winner was too small compared with the average loser. I do not think the entry side gets cleared completely, but the exit side is where the damage shows up first.GateGrid AI and MLScore GF-T4 GB both finished at a 50.0% win rate. But GateGrid’s two wins totaled only +93 yen, while its two losses totaled -796 yen. MLScore looked slightly healthier on winners, with +351 yen of gross profit, but the -600 yen loss hit like the kind of print you do not just scroll past.The design notes for GateGrid say the bot is built to leave a reason trail, including messages like AI_SKIP(sess=NY gate=0.50 base_thr=0.54 adj_thr=0.55) and OLLAMA_HOLD. That is exactly the kind of log I wanted beside today’s trades. The MT5 export tells me where the damage happened, but not enough about why the model stayed in, closed there, or failed to avoid the bad sequence.GateGrid AIGateGrid AI was the heaviest drag of the day at -703 yen. The first two exits were small wins, +85 yen and +8 yen, which is fine on paper but too small to matter once the later pair of losses arrived. When I saw -364 yen and -432 yen printed back to back, the earlier +93 yen stopped feeling like progress.This is a payoff problem before it is a win-rate problem. A 50.0% win rate can work if the average win and average loss are balanced, but here the payoff ratio was only 0.12. The grid did take profit when it had a chance, but the losing leg widened far beyond the winning leg. My suspicion is exit timing or trailing behavior, not the mere fact that it entered.The architecture is meant to filter entries through CatBoost and then use Ollama for situational judgment. That design is still attractive, because a bot that can say OLLAMA_HOLD has at least some mechanism for refusing trades. Today’s report, though, only shows the final broker-side outcome. Without the actual decision log, I cannot tell whether the AI wanted to hold, whether the trailing logic waited too long, or whether the grid structure simply accepted too much downside.BoundSniperBoundSniper was the only realized winner, but only by +3 yen. That number almost feels like a joke, not because it is bad, but because it survived while the more complex bots took the larger hits. The first close had a loss plus swap cost, then the next two closes recovered enough to end slightly positive.This bot is not trying to be smart in the same way. It passes TradingView signals to MT5, so the performance review belongs mostly to the upstream strategy and execution reliability. Still, the payoff ratio was 0.51, better than the day’s total, and the largest loss was contained at -159 yen. That boring containment mattered.The open USDJPY- short was still floating at -30 yen at the report cutoff. I did not include that in realized P/L, but I would not ignore it either. BoundSniper’s job is clean execution, and today it did that well enough to avoid becoming the story.MLScore GF-T4 GBMLScore GF-T4 GB had the day’s largest single realized loss at -600 yen. That one trade shaped the whole read. The bot had two winners, +151 yen and +200 yen, so it was not failing every time it touched the market. Still, one oversized loss swallowed both winners and left the bot at -487 yen.This is where the “50% win rate” line becomes almost misleading. Two wins and two losses can be perfectly acceptable, but only if the losses are not three times the wins. The payoff ratio was 0.42, which is not hopeless, yet the max loss was too large relative to the rest of the day.The report also showed an open GBPJPY- sell with -124 yen unrealized at cutoff. That makes the exit question harder to ignore. The closed book already had one oversized stop, and the open book was still underwater. I would want to inspect whether the stop width is static, volatility-adjusted, or simply too generous for the current GBPJPY movement.LLMBridgeTraderLLMBridgeTrader ended at -258 yen, with the same 2W / 2L pattern. The difference is that the losses were smaller than MLScore’s and GateGrid’s largest hits, but the wins were also small. A payoff ratio of 0.34 is still too thin.The bot design says it can return not only BUY / SELL / NONE, but also OPEN, HOLD, CLOSE, and REVERSE, along with confidence, setup type, SL pips, TP pips, and entry or exit reasons. That is the part I most want to read on a day like this. The MT5 comments only preserved LLMBridgeTrader_, so I can see the executed trades, but not the reasoning that led to closing at -144 yen or taking the later +77 yen and +58 yen.For an AI-led bot, the entry is only half the experiment. The more interesting question is whether the model knows when its original idea has expired. Today’s LLMBridge result was not catastrophic, but it still leaned toward small winners and larger losers. That pattern needs a stronger close-or-hold audit.SummaryThe day did not collapse because every bot was wrong. It collapsed because the losing trades were allowed to become too heavy relative to the winners. BoundSniper, the least “AI-planner” style bot, was the only one to end positive, and that is a little uncomfortable.The next review should focus less on signal accuracy and more on the moment each bot decides that a trade is no longer worth holding. That is where the money left the account today.Editor’s note: next time, paste the actual AI decision logs too. The numbers are enough to score the day, but the logs are what turn it into a real LLM trading experiment. This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit fxaibotlab.substack.com/subscribe