Looking back at the operating logs of the AI trading bots on July 15 and 16, a clear and somewhat painful theme has emerged. It is the reality that “the timing of the exit (closing a position) is far more important than the frequency or accuracy of the entry.”
The Illusion of Win Rate and the Payoff Ratio (Lessons from July 15)
On July 15, we ran four MT5 bots, ending the day with a total loss of -1,456 JPY.
Some bots recorded a 50% win rate that day, yet the account balance still decreased. Overall, the record was 6 wins and 8 losses (a 42.9% win rate), meaning they were not losing excessively. The bots were not wildly misjudging the market direction all day.
The problem was that they were “paying too much when they were wrong.” The payoff ratio was only 0.51, meaning the average profit was only half the size of the average loss. Even if small profits are accumulated, a single loss, inflated by a delayed exit, can wipe them all out. Faced with such results, the “win rate” metric feels like mere noise.
AI’s Struggle and the Victory of Simplicity (Lessons from July 16)
On the following day, July 16, we ran five bots, including the newly added MAribbonTrader. The realized profit and loss was -464 JPY.
On this day, the difference in the bots’ “approaches” clearly divided the results. The most outstanding performer was, ironically, the least “conversational” bot, ML_ScoreAnalyst. Without hesitation, this system successfully hit the take-profit on both of its two trades, generating a profit of +502 JPY on its own.
In contrast, the bots that relied on LLMs (Large Language Models) or image recognition AI for judgment struggled with “cutting losses,” “switching positions,” and “staying out of the market.” The decision of whether to “HOLD” or “CLOSE” a position after entering proved to be the most costly challenge.
Behavior and Challenges of Each Bot
Over these two days, the different risks associated with each of the five bots became apparent.
* GateGrid AI
On the 15th, despite a 50% win rate, delayed exits made the trades too costly, resulting in a loss. On the 16th, it showed inexplicable behavior, closing a trade in the same second it was opened, resulting in a -13 JPY loss. While the entry filter is functioning, a detailed log review of the handoff between execution and closing is necessary.
* BoundSniper Bot
This is a rule-based bot that executes signals received from TradingView. It had a 50% win rate on the 15th, but on the 16th, it suffered seven consecutive losses after one win, dropping its win rate to 12.5%. The issue is not the execution engine itself; improving the quality of the upstream signals and revising the exit rules during losing streaks is urgently needed.
* LLMBridgeTrader
On the 15th, despite a 33.3% win rate, it limited its losses and showed a decent structure with a payoff ratio of 1.23. However, on the 16th, immediately after a loss, it quickly took a position in the opposite direction and held an unrealized loss (-145 JPY). Measures are needed to prevent delayed withdrawals and overreaction (chasing) after a loss.
* ML_ScoreAnalyst (formerly MLScore GF-T4)
On the 16th, it successfully took profit on two GBPJPY long trades, achieving a commendable result of +502 JPY. However, on the 15th, it had a record of “a winning trade of +300 JPY and a losing trade of -599 JPY,” revealing the issue of its stop-loss being too wide. We must continue to evaluate the “true cost” of this wide stop-loss, rather than just trusting clean wins like those on the 16th.
* MAribbonTrader
This bot, which uses image recognition AI to read charts, held a position for over 12 hours on the 16th, resulting in the largest single loss of the day (-447 JPY). The market context must have changed during that long holding period, making it necessary to verify whether the exit mechanism is functioning properly.
Next Steps: Towards Exit Discipline
In live automated trading, what protects the day’s profits is the decision that “I will not hold this position any longer.” The results of these tests teach us that future system improvements should focus on “exit discipline” rather than “entry confidence.”
In the next iteration, we will implement the following changes:
* Detailed State Logging: Record the position direction, unrealized P/L, model confidence, and the reason for the decision (e.g., HOLD or CLOSE) as a compact log to make the validity of the closing decision verifiable.
* Review Maximum Holding Time: Set re-evaluation triggers based on elapsed time to reduce the risks associated with holding positions for a long time.
* Introduce Cooldowns: Prevent immediate entries right after a loss is finalized, separating genuine trend reversals from “emotional chasing.”
Systems equipped with advanced AI models have, ironically, left us with the most human and difficult question of “when to exit.” However, these clear differences between the systems and the records of their failures are precisely the most valuable data we need to push them to the next level.
This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit fxaibotlab.substack.com
