A 73.9% Win Rate Still Lost Money: The Week I Stopped Trusting Entries and Started Rebuilding Exits

The headline from this week is uncomfortable, but useful: the bots were not simply bad at entries. Some of them were right more often than they were wrong, and that is exactly what made the result harder to ignore. Across the June 22–26 run, the total came in at -3,333 yen, and the number that bothered me most was not the loss itself. It was seeing a bot clear a 70% win rate and still fail to protect the account.This week pushed the experiment away from “Can the AI predict the next move?” and closer to “Can the system stop holding a broken idea?” That sounds like a small wording change, but in live MT5 operation it changes almost everything. Entry logic is visible and satisfying; exit discipline is less glamorous, and it is also where the damage was hiding.Bot-by-Bot Results■ GateGrid AIPeriod: June 22–26 weekly reviewRecord: 17W / 6L on June 24 reference day (Win rate 73.9%)Net P/L: Negative for the reviewed periodGross profit: Not fully disclosed in this logGross loss: Not fully disclosed in this logPayoff ratio: Weak, due to outsized lossesMax loss: Around -1,100 yen referenced■ MLScore GF-T4 GBPeriod: June 22–26 weekly reviewRecord: 1W / 1L on one reviewed day (Win rate 50.0%)Net P/L: Negative pressure on the weekly totalGross profit: +140 yen referencedGross loss: Around -600 yen referencedPayoff ratio: About 0.23 from the referenced pairMax loss: Around -600 yen referenced■ LLMBridgeTraderPeriod: June 22–26 weekly reviewRecord: 4W / 3L (Win rate 57.1%)Net P/L: Positive contributionGross profit: Not fully disclosed in this logGross loss: Not fully disclosed in this logPayoff ratio: 2.43Max loss: Not disclosed in this log■ TotalPeriod: June 22–26Record: Mixed across botsNet P/L: -3,333 yenGross profit: Not fully disclosed in this logGross loss: Not fully disclosed in this logPayoff ratio: Mixed, with LLMBridgeTrader offset by GateGrid AI and MLScore GF-T4 GBMax loss: Around -1,100 yen referencedToday’s Theme: The Trap Was Not Low AccuracyThe obvious story would be “the bots lost because the AI was wrong.” That is too simple, and honestly it does not match the logs. GateGrid AI, for example, produced a 17W / 6L day with a 73.9% win rate. When I saw that number next to a negative result, I had to pause for a second. A system that wins that often should not feel that fragile.The real issue was payoff structure. Small wins were being collected, then one large loss came in and erased the quiet work before it. A -700 yen or -1,100 yen stop on a grid-style bot is not just another losing trade; it is a design warning. The bot was not failing every minute. It was failing at the exact moment when the position idea had already been invalidated.GateGrid AI: The Grid Needed a Harder LineGateGrid AI is built around more than a simple grid. It uses CatBoost-style entry filtering, local LLM judgment through Ollama, ATR checks, session filters, spread monitoring, and adaptive grid management. On paper, that gives the bot several chances to avoid weak trades. In practice, the week showed that avoiding bad entries is not enough when a grid has already started stacking exposure.The painful part was the familiar one: many small wins, then one oversized loss. The bot could be “right” most of the time and still let one grid collapse dominate the week. I do not want to overstate certainty here, but the failure point looks more like the exit than the entry. The bot needed a rule that says, “This trade idea is no longer alive,” instead of letting the grid structure argue for more patience.The new rule is a forced exit after the second grid position is formed. If price breaks back through the first entry line and then continues a defined number of pips against the position, the system cuts. That condition matters because it is not random noise anymore. The second layer is already in, the first level has been violated, and the market has kept moving the wrong way. At that point, letting the position breathe may just be another word for postponing the loss.MLScore GF-T4 GB: Breakout Risk Had to Be FixedMLScore GF-T4 GB had a different problem. Even on a 1W / 1L sample, the structure was ugly: around +140 yen on the win and around -600 yen on the loss. That payoff ratio, roughly 0.23, is the sort of number that makes a 50% win rate almost irrelevant. I saw the +140 yen and -600 yen pairing and thought, not again — not because the trade lost, but because the ratio had already decided the result.The update here is simpler and more mechanical. For breakout setups, TP is now fixed at 30 pips and SL at 25 pips. Range logic stays unchanged, because the behavior of a range setup is different. But breakout trades should either accelerate or fail quickly, and the old structure allowed too much room for a failed breakout to become a large wound.By forcing the risk-reward to about 1.2, the bot is no longer allowed to take a breakout just because the score looks good. The AI or model can still identify the setup, but the system now refuses to let conviction stretch the stop too far. That is less romantic than letting the bot adapt freely, but live trading has a way of punishing freedom when it is not boxed in.LLMBridgeTrader: Lower Win Rate, Better Damage ControlLLMBridgeTrader was the useful counterexample. With 4W / 3L and a 57.1% win rate, it did not look like the cleanest bot by accuracy. Yet its payoff ratio was 2.43, and that changed the week’s interpretation. A lower win rate with better exits can beat a high win rate with oversized losses.This bot is closer to the experiment I actually want to run: not just asking the LLM for BUY, SELL, or NONE, but letting it reason about position actions such as OPEN, HOLD, CLOSE, and REVERSE. That added responsibility is risky, especially when the model can overreact or narrate confidence too well. Still, the week suggested that the exit side is where LLM judgment may be most interesting. The model does not need to be right all the time if it can stop being wrong quickly.I would not call this solved. REVERSE logic still needs caution, confidence thresholds need more testing, and the live-vs-backtest gap is always waiting in the background. But among the bots this week, LLMBridgeTrader showed the cleanest relationship between being wrong and paying a reasonable price for it.SummaryThe week ended at -3,333 yen, but the useful part was not the amount. The useful part was the pattern: high win rate did not save a weak exit design, and a modest win rate looked far healthier when the losses were contained.So the next phase is not more prediction for its own sake. GateGrid AI now has a forced retreat rule for grid breakdowns. MLScore GF-T4 GB has fixed breakout risk with TP 30 / SL 25. LLMBridgeTrader remains the experiment in whether an LLM can handle not only entries, but the harder question of when to stop believing its own plan.I still like AI-driven trading systems. I just trust them less when they are only good at starting trades. 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