AI FX Bot Lab: Real Trading Experiments

Kimi | Japan FX Bot Lab

0,0 (0)
Technisch nieuws
Maandelijks bijgewerkt

Can AI really trade forex? AI FX Bot Lab is a real-time experiment from Japan, where I build and test AI-assisted FX trading bots using MT5, Python, machine learning, and local LLM tools. I share live results, failures, risk lessons, and bot improvements from rule-based, AI-driven, and ML + LLM hybrid systems. Not financial advice. fxaibotlab.substack.com

1 dag geleden

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 referenced Today’s Theme: The Trap Was Not Low Accuracy The 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 Line GateGrid 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 Fixed MLScore 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 Control LLMBridgeTrader 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. Summary The 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 would like to discuss this with other subscribers or get access to bonus episodes, visit fxaibotlab.substack.com

2 min.
2 dgn geleden

The 57.1% Bot Carried the Day, Not the Perfect One

The strongest bot today was not the one with a 100% win rate. BoundSniper closed its single trade cleanly for +10 yen, which is fine, but the real driver was LLMBridgeTrader: 4 wins, 3 losses, +517 yen realized, and a payoff ratio of 2.43. I had to look twice at that combination because 57.1% does not sound dominant until the average win starts doing the work. The full realized result across the four bots was +524 yen. That number is not huge by itself, but the shape matters more than the size today. LLMBridgeTrader absorbed three losing trades, including a -194 yen hit that made me pause for a second, and still finished well ahead because its winners were allowed to breathe. That is exactly the kind of exit behavior I wanted to watch in this experiment. For the bot roles, I am treating BoundSniper as the TradingView execution bot, LLMBridgeTrader as the AI-led position planner, and GateGrid AI as the CatBoost plus Ollama gated grid system, based on the saved bot memo. ts ■ GateGrid AI +71 yenRecord: 1W / 1LWin rate: 50.0%Gross profit: +89 yenGross loss: -18 yenPayoff ratio: 4.94Max loss: -18 yen ■ BoundSniper +10 yenRecord: 1W / 0LWin rate: 100.0%Gross profit: +10 yenGross loss: 0 yenPayoff ratio: N/A, no losing tradeMax loss: 0 yen ■ LLMBridgeTrader +517 yenRecord: 4W / 3LWin rate: 57.1%Gross profit: +748 yenGross loss: -231 yenPayoff ratio: 2.43Max loss: -194 yen ■ MLScore GF-T4 GB -74 yenRecord: 1W / 1LWin rate: 50.0%Gross profit: +202 yenGross loss: -276 yenPayoff ratio: 0.73Max loss: -276 yen ■ Total +524 yenRecord: 7W / 5LWin rate: 58.3%Gross profit: +1,049 yenGross loss: -525 yenPayoff ratio: 1.43Max loss: -276 yen Open positions were still on the board at the report cutoff. LLMBridgeTrader had an unrealized +92 yen position, while MLScore GF-T4 GB had an unrealized -180 yen position. So the realized result was +524 yen, but the mark-to-market feel of the day was closer to +436 yen. I do not want to blur those two numbers, because open P/L can turn into a very different story by the next report. Today’s theme Today was an exit test more than an entry test. The entries mattered, of course, but the day was decided by what each bot did after being in the trade: whether it cut too early, held long enough, or let one loss dominate the session. That is especially important for the LLM-driven bots. When the model is allowed to decide not just direction but also OPEN, HOLD, CLOSE, or REVERSE, the question changes. I am no longer only asking, “Did it predict the next move?” I am asking whether it knew when to stop insisting on its first idea. GateGrid AI GateGrid AI only closed two positions today, one win and one small loss. The result was +71 yen with a 50.0% win rate, but the payoff ratio was 4.94 because the losing trade was only -18 yen. That -18 yen loss is the kind of small scratch I can live with; it does not force the next trade to become a rescue mission. The interesting part is that GateGrid AI did not need many trades to stay positive. This bot is built around filtering, with CatBoost first narrowing the entry probability and Ollama acting as a second gate. On a day like this, the low trade count is not automatically a weakness. It might simply mean the bot found only a couple of situations worth touching. The exit also looked controlled. There was no oversized loss hiding under a good win rate, and no open position left behind at the cutoff. I would not call this a strong day, but it was a clean one. For a grid-style bot, clean can be more valuable than exciting. BoundSniper BoundSniper had the cleanest record on paper: 1 win, 0 losses, +10 yen. A 100.0% win rate always looks nice for a second, then the amount pulls it back to earth. This bot did its job as an execution layer, and that is probably the correct way to read the result. Because BoundSniper is not trying to be an AI trader, I do not want to over-interpret the trade. It received the TradingView-side signal, entered, exited, and ended positive. No drama, no open exposure, no large adverse move. The limitation is that one trade tells us almost nothing about edge. It tells us the pipeline worked. That matters, especially in live automation, but the performance story belongs somewhere else today. LLMBridgeTrader LLMBridgeTrader was the center of the day. It closed 7 trades, won 4, lost 3, and still ended at +517 yen. The payoff ratio of 2.43 is the key number here. A 57.1% win rate with that payoff profile can survive a few mistakes, and today it did exactly that. The bot took three losses: -18 yen, -194 yen, and -19 yen. The -194 yen loss bothered me because it was large enough to test whether the rest of the session would become damage control. But the later winners, especially +369 yen, changed the whole texture of the result. That trade is where the bot stopped looking merely active and started looking useful. The open position also matters. At the cutoff, LLMBridgeTrader was holding a EURUSD sell position with +92 yen unrealized. That suggests the bot had not simply churned itself flat after the realized gain. It was still holding a live idea. Whether that was discipline or stubbornness will only be clear after the next close, but for today the exit logic looked better than I expected. MLScore GF-T4 GB MLScore GF-T4 GB was the weak spot. It had one win of +202 yen and one loss of -276 yen, leaving the realized result at -74 yen. The win rate was 50.0%, the same as GateGrid AI, but the payoff ratio was only 0.73. Same win rate, completely different feel. The maximum loss was -276 yen, which was also the largest closed loss across all bots. That is the kind of number that changes how I look at a flat-looking record. One win and one loss should be almost boring, yet here the loss carried more weight than the win. There was also an open GBPJPY buy position with -180 yen unrealized at the report cutoff. That does not mean the trade is wrong, but it does mean the day was not really finished for this bot. My suspicion is that the issue is not entry alone. The exit line, or maybe the distance between “hold” and “admit defeat,” needs more review. Summary Today’s realized total was positive, but the real lesson came from the contrast between win rate and loss shape. BoundSniper had the perfect record and only added +10 yen. GateGrid AI won only half its trades and still stayed clean. LLMBridgeTrader carried the account because its average winner was large enough to cover its misses. MLScore GF-T4 GB reminded me that a 50.0% day can still feel heavy when the bigger side is the loss. I am not ready to call LLMBridgeTrader stable from one good session. But today, the AI-led exit decisions looked less like noise and more like something worth continuing to measure. ② Substack Note The surprise from June 26 was simple: the 100% win-rate bot was not the star. BoundSniper went 1-for-1 and made +10 yen. Clean, but tiny. LLMBridgeTrader went only 4W / 3L, yet finished +517 yen because its winners were much larger than its losers. The payoff ratio came in at 2.43, and that changed the whole day. Total realized P/L across the four MT5 bots: +524 yen. The experiment is becoming less about “can the LLM pick direction?” and more about “can it stop holding the wrong idea, while staying long enough with the right one?” 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

2 min.
3 dgn geleden

A ¥728 Exit Turned the Whole Bot Run Negative

Conclusion The day ended at -1,180 yen across four MT5 bots, and the uncomfortable part is that the record did not look broken at first glance. There were 9 winning exits and 7 losing exits overall, so the surface was not ugly. But the payoff ratio was only 0.26, and the largest single loss was -728 yen from GateGrid AI. I paused on that number for a moment, because it was bigger than the total gross profit of every bot combined. The main theme today is exit quality. For the bots that hand judgment to an LLM or an AI layer, the question is no longer just “was the entry right?” It is whether the model knows when the original idea has expired. Today, that part still feels unfinished. Bot-by-bot results ■ GateGrid AI -512 yen Record: 5W / 2L Win rate: 71.4% Gross profit: +235 yen Gross loss: -747 yen Payoff ratio: 0.13 Max loss: -728 yen ■ BoundSniper -25 yen Record: 2W / 1L Win rate: 66.7% Gross profit: +52 yen Gross loss: -77 yen Payoff ratio: 0.34 Max loss: -77 yen ■ LLMBridgeTrader -175 yen Record: 1W / 3L Win rate: 25.0% Gross profit: +162 yen Gross loss: -337 yen Payoff ratio: 1.44 Max loss: -201 yen ■ MLScore GF-T4 GB -468 yen Record: 1W / 1L Win rate: 50.0% Gross profit: +144 yen Gross loss: -612 yen Payoff ratio: 0.24 Max loss: -612 yen ■ Total -1,180 yen Record: 9W / 7L Win rate: 56.3% Gross profit: +593 yen Gross loss: -1,773 yen Payoff ratio: 0.26 Max loss: -728 yen Today’s theme The strange thing about this run is that the losing day was not caused by constant bad entries. GateGrid AI won most of its exits. BoundSniper also had more winning closes than losing ones. Even MLScore was split one and one. And still, the day sank because the losing trades were far larger than the winning trades. That makes this less of a signal problem and more of an exit problem. The bots can find small profitable windows, but when price moves against them, the stop behavior and close timing are still too heavy. A 71.4% record with a 0.13 payoff ratio is not strength; it is a warning label written in small numbers. GateGrid AI GateGrid AI was the most painful bot to read today. It finished with 5 wins and 2 losses, which sounds fine until the -728 yen loss appears at the end. The five wins added only +235 yen, so one late loss erased all of them and then some. Seeing -728 yen after a string of small wins had that familiar “not this shape again” feeling. The entry filter may still be doing something useful. GateGrid AI did not spray random losing trades all day. The problem is that the grid logic and exit logic allowed one position to become too large relative to the normal win size. If the bot is designed to collect small moves, then a single loss cannot be allowed to equal fifteen small wins. That is where the current structure looks fragile. For an ML plus LLM hybrid bot, the next review should focus on the moment it stops believing in the setup. CatBoost and Ollama may help filter entries, but once a position is live, the bot also needs a stronger “the idea is no longer valid” trigger. I suspect the issue is not the first decision. It is the delay in giving up. BoundSniper BoundSniper ended at -25 yen, and this one is a different kind of result. The trade logic itself is not AI-driven; it passes TradingView signals into MT5. So I do not read this as a model judgment failure. It is more about execution, signal timing, and the cost carried by the position. The first close showed +28 yen on price movement, but after swap it became a net drag. That is small, but it matters because the other wins were only +10 yen and +42 yen. A tiny edge disappears quickly when the holding cost is not small relative to the expected win. BoundSniper did not collapse today. Still, its payoff ratio was only 0.34 on a net basis. That means it needs either cleaner exits or larger average wins, because a bot that depends on TradingView rules cannot count on AI interpretation to rescue weak trade economics later. LLMBridgeTrader LLMBridgeTrader is the most interesting bot today, even though the result was negative. It only won once and lost three times, but its payoff ratio was 1.44. That means the structure is not hopeless. One winning exit was large enough to cover more than one average loss, at least in theory. The issue is frequency and sequence. After the +162 yen win, the bot took -57 yen, then -201 yen, then -79 yen. The model is allowed to decide OPEN, HOLD, CLOSE, and REVERSE, so the exit decision is part of the experiment, not just a mechanical afterthought. Today, the AI did close trades, but it did not avoid the cluster of small-to-medium losses that followed. This is where LLM trading gets uncomfortable. The model can describe a reason, and the log can preserve that reason, but the account only cares whether the reason led to a better exit. I would not throw away this setup from one day. I would look harder at confidence thresholds for CLOSE and REVERSE, because the bot may need to be more conservative once it has already taken a directional loss. MLScore GF-T4 GB MLScore GF-T4 GB had only two closed results, so I do not want to overstate the sample. Still, the shape was clear: one net win of +144 yen after swap, then one loss of -612 yen. That gave it a 50.0% record but a payoff ratio of only 0.24. Half right is not enough when the wrong side is four times heavier. The -612 yen loss is the second biggest single loss of the day. It did not come from a long sequence of mistakes; it came from one trade that carried too much damage. That makes the review simple, though not easy. The bot needs a better hard stop, or it needs to size down when the expected stop distance is wide. This bot may still be useful as a scoring layer, but today it behaved like a model that can be directionally right sometimes while still failing the risk shape. I do not have enough from one day to say the score is bad. The exit width is the part I would question first. Summary Today was not a clean “AI failed” day. It was more specific than that. The bots found winners, and some of the entry logic looked alive, but the loss distribution was badly tilted. Total gross profit was +593 yen against -1,773 yen in gross losses, and that gap tells the story more honestly than the win count. For the next tuning pass, I would not start by chasing more entries. I would start with maximum loss rules, earlier invalidation, and stricter handling of HOLD turning into CLOSE. The experiment is still worth running, but today the market reminded me that a smart entry is only half a trade 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

2 min.
5 dgn geleden

71.4% Win Rate Still Lost Money: MT5 LLM Bot Log for June 23, 2026

The total result was positive: +455 yen across 16 closed trades. That sounds clean enough, and honestly, seeing the total stay above zero was a relief. But the win rate is almost the least interesting number today. The whole portfolio went 13W / 3L, yet the total payoff ratio was only 0.54, meaning the average losing trade was still larger than the average winning trade. The main issue was not entry accuracy. It was loss shape. GBPUSD, which I track as GateGrid AI based on the bot setup notes, won 5 out of 7 trades and still finished at -132 yen. That number made me pause a little, because this is the exact kind of day where a “good win rate” can hide a weak exit structure. The bot notes describe GateGrid AI as a CatBoost + Ollama multi-gate system, with logs such as AI_SKIP(sess=NY gate=0.50 base_thr=0.54 adj_thr=0.55) and OLLAMA_HOLD, so the design is already focused on filtering bad entries. Today’s P/L suggests the next place to inspect is after entry: when to stop holding, when to cut, and whether the grid loss is allowed to grow too far. Bot Results ■ GBPJPY Bot +200 yenRecord: 1W / 0LWin rate: 100.0%Gross profit: +200 yenGross loss: 0 yenPayoff ratio: N/AMax loss: 0 yen ■ LLMBridgeTrader +167 yenRecord: 3W / 1LWin rate: 75.0%Gross profit: +173 yenGross loss: -6 yenPayoff ratio: 9.61Max loss: -6 yen ■ BoundSniper Bot +220 yenRecord: 4W / 0LWin rate: 100.0%Gross profit: +220 yenGross loss: 0 yenPayoff ratio: N/AMax loss: 0 yen ■ GateGrid AI -132 yenRecord: 5W / 2LWin rate: 71.4%Gross profit: +207 yenGross loss: -339 yenPayoff ratio: 0.24Max loss: -205 yen ■ Total +455 yenRecord: 13W / 3LWin rate: 81.3%Gross profit: +800 yenGross loss: -345 yenPayoff ratio: 0.54Max loss: -205 yen Today’s Theme: Loss Size Beat Win Rate The cleanest bot on paper was BoundSniper Bot on USDJPY. Four trades, four wins, +220 yen. No losing trade means I cannot evaluate the payoff ratio yet, but as a pure execution bot that follows TradingView-side signals, this was a very good session. LLMBridgeTrader on EURUSD was the strongest from a risk-shape view. It took one tiny loss of -6 yen and then built +167 yen total. A payoff ratio of 9.61 is almost too clean for one day, so I would not overtrust it yet, but the exit behavior looked good. It did not let the bad trade become a story. GBPJPY Bot had one trade and closed +200 yen. That result helps the day, but one trade is too thin to judge. Still, one clean winner with no damage is not something I complain about. GateGrid AI is where the day gets interesting. Five wins created only +207 yen, while two losses took -339 yen. The average win was 41.4 yen, while the average loss was 169.5 yen. That means this bot needs roughly an 80% win rate just to break even under today’s loss shape. A 71.4% win rate sounds good until the math quietly turns against it. Bot-by-Bot Read BoundSniper Bot did what a rule-following execution bot is supposed to do: it captured small USDJPY moves without taking damage. Since this bot itself does not make the market prediction, the key review point is not “AI judgment,” but signal quality and execution slippage. Today, nothing in the result suggests an execution problem. LLMBridgeTrader had the best balance. A -6 yen loss is the kind of loss I like to see from an AI-led bot because it means the system was willing to abandon the idea quickly. The +129 yen EURUSD short was the main contributor, and the later +36 yen and +8 yen trades added without giving much back. GateGrid AI needs the most review. The entry filter may still be useful, but the exit side looks loose. I cannot say the AI made a bad judgment without the matching daily log, but the numbers point in that direction: once the position was allowed, the losing side stayed open long enough to erase five smaller wins. The -205 yen loss was the trade that stopped me. Wrap-Up Today ended positive, but not because every bot was healthy. The portfolio survived because USDJPY, GBPJPY, and EURUSD covered the GBPUSD damage. For the next review, I would not start by improving win rate. I would start with GateGrid AI’s maximum loss, trailing behavior, and the exact moment it chose not to exit. A profitable day can still leave homework. 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

2 min.
6 dgn geleden

The Average Loss Beat the Whole Bot Lineup

Today was not a day I want to judge by win count. The four-Bot portfolio closed at -2,457 yen, and the reason was not that every system failed at entries. It was simpler and more uncomfortable: the losing trades were too heavy compared with the winners. The cleanest Bot on the sheet was LLMBridgeTrader, which finished at +353 yen with a payoff ratio of 5.24. That number stopped me for a second, because it was the only Bot that looked like it knew how to be wrong cheaply. GateGrid AI, on the other hand, had many small profitable exits but ended at -1,482 yen because one loss, -733 yen, swallowed too much of the day. The cumulative result is therefore -2,457 yen, from a combined starting balance of 168,415 yen to 165,958 yen across the four accounts. Bot-by-bot results ■ GateGrid AI -1,482 yenPair: GBPUSDClosed trades: 14Record: 10W / 4LWin rate: 71.4%Gross profit: +519 yenGross loss: -2,001 yenPayoff ratio: 0.10Max loss: -733 yen ■ MLScore GF-T4 GB -244 yenPair: GBPJPYClosed trades: 1Record: 0W / 1LWin rate: 0.0%Gross profit: 0 yenGross loss: -244 yenPayoff ratio: N/AMax loss: -244 yen ■ LLMBridgeTrader +353 yenPair: EURUSDClosed trades: 4Record: 3W / 1LWin rate: 75.0%Gross profit: +377 yenGross loss: -24 yenPayoff ratio: 5.24Max loss: -24 yen ■ BoundSniper Bot -1,084 yenPair: USDJPYClosed trades: 6Record: 3W / 3LWin rate: 50.0%Gross profit: +216 yenGross loss: -1,300 yenPayoff ratio: 0.17Max loss: -468 yen ■ Total -2,457 yenClosed trades: 25Record: 16W / 9LWin rate: 64.0%Gross profit: +1,112 yenGross loss: -3,569 yenPayoff ratio: 0.18Max loss: -733 yen Today’s theme: the exit mattered more than the entry The main story is not “which Bot had more winning trades.” It is how much damage each Bot allowed when the trade was wrong. Across the portfolio, the average winning trade was about 69.5 yen, while the average losing trade was about 396.6 yen. That gap is too wide. It means one loss needed almost six average wins just to repair it. This is where LLMBridgeTrader stood out. Its one losing trade was only -24 yen, and the three winners were large enough to cover it without drama. I do not want to overpraise one day of data, but this is the shape I want from an AI-driven trading Bot: not perfect prediction, just fast retreat when the premise weakens. The MT5 report did not include the full AI decision logs for this day. I only had execution comments such as “LLMBridgeTrader_”, “[sl 1.14622]”, “[sl 213.719]”, and “BoundSniper OPEN”. That matters, because the most useful analysis would connect the Bot’s actual reasoning to the exit result, not just the final yen amount. GateGrid AI: many small exits, one large wound GateGrid AI is supposed to filter entries with a multi-stage design: model gate, Ollama-style judgment, volatility checks, session control, and grid management. On paper, that should protect it from bad conditions. Today’s results still show a familiar grid problem: the winners were small, and the losses had room to grow. The gross profit was +519 yen, but gross loss reached -2,001 yen. The payoff ratio was 0.10, which is the uncomfortable part. A small winner like +8 yen or +19 yen feels harmless while it is happening, but it does not build enough cushion when a -733 yen exit appears later. Seeing that -733 yen line made me pause, because this was not just a losing trade; it was a statement about the exit width. The issue is probably not entry frequency alone. The question is whether the Bot should cut the grid earlier when price keeps moving against the cluster. I still need the actual AI_SKIP / OLLAMA_HOLD / close-reason logs to say that with confidence, but the PnL shape points toward exit control. MLScore GF-T4 GB: one trade, no room to judge MLScore GF-T4 GB had only one closed trade on GBPJPY and finished at -244 yen. The MT5 comment shows “[sl 213.719]”, so this was a stop-based exit rather than an active recovery sequence. There is not enough here to judge the model. One trade can be noise. Still, for a portfolio day, a single stop loss matters when the rest of the Bots are already carrying wide downside. I would treat this Bot as “not guilty yet,” but not invisible either. LLMBridgeTrader: the best result came from losing small LLMBridgeTrader was the one clean positive result: +353 yen, with only -24 yen of gross loss. That is the part I care about more than the win count. It did not need many trades to recover; the negative trade was simply small enough. The MT5 report shows two profitable exits marked with stop-style comments, including “[sl 1.14622]” and “[sl 1.14192]”. That looks like a protective stop or locked-in exit behavior rather than a raw fixed loss. If that reading is right, then the Bot’s exit design did more work than the entry direction. This is the sort of behavior I want to keep watching. LLMBridgeTrader is the Bot where the AI is meant to decide not only BUY / SELL, but also OPEN / HOLD / CLOSE / REVERSE. Today, I cannot see the actual reasoning text, but the result suggests that the exit side deserves more credit than the entry side. BoundSniper Bot: the executor did its job, the loss size did not BoundSniper Bot finished at -1,084 yen on USDJPY. This Bot is not trying to predict the market by itself. It receives TradingView signals through the webhook route and sends the corresponding orders to MT5, so the real evaluation belongs upstream: signal quality, stop size, and whether the exit logic is too late. The shape was rough. Gross profit was only +216 yen, while gross loss reached -1,300 yen. The largest loss was -468 yen, and there were three losses in that same heavy zone: -420, -412, and -468. It is hard not to see that as a structural issue rather than a bad tick. BoundSniper may be doing exactly what it was told to do. That does not make the strategy healthy. For this Bot, I would not start by changing the MT5 bridge; I would first review the TradingView exit condition and the distance between “wrong” and “closed.” Summary The portfolio did not lose because every Bot was directionally bad. It lost because the negative trades were allowed to become too large compared with the average positive trade. LLMBridgeTrader was the exception, and that is why it is the main reference point for the next review. For the next run, I want to see the actual AI decision logs beside the trades. The yen result tells me what happened; the logs would tell me whether the Bot hesitated, protected, reversed, or simply waited too long. Editing note to myself: next time, paste the AI reasoning logs too, especially OPEN / HOLD / CLOSE / REVERSE reasons, confidence, setup type, and close reason. 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

2 min.
21 jun

Win Rate Lied This Week: The Rule-Based Bot Survived While the LLM Bots Bled

Win Rate Lied This Week: The Rule-Based Bot Survived While the LLM Bots Bled MT5 LLM auto-trading report, June 15–19, 2026.This is Day 5 of the running log. The four-bot portfolio finished the window at -3,490 yen cumulative, even though several days showed decent-looking win rates on the surface. The uncomfortable part is not the loss itself. It is the shape of the loss. GateGrid AI kept showing high win-rate behavior, yet one large basket loss could erase a pile of small wins. BoundSniper, the least “AI-like” bot in the group, quietly ended as the only steady positive contributor. That made me pause a bit; it is not the result I wanted from an LLM-heavy experiment, but it is the result on the screen. The five-day total moved like this: +282 yen, -1,445 yen, -1,615 yen, -464 yen, -248 yen. Seeing a 53.3% win-rate day end at -1,445 yen still feels wrong at first glance. Then the payoff ratio explains it. The system was winning often enough, but not winning large enough. Bot-by-bot results Trade-level counts were not fully included in the provided block, so I treated each date’s final Bot result as one result unit. Where specific trade-level clues were given, I mention them in the analysis rather than pretending the missing rows are available. ■ GateGrid AI -1,367 yenRecord: 2 positive days / 3 negative daysDay-level win rate: 40.0%Gross profit: +203 yenGross loss: -1,570 yenPayoff ratio: 0.19Max reported daily loss: -712 yen ■ BoundSniper +271 yenRecord: 4 positive days / 1 negative dayDay-level win rate: 80.0%Gross profit: +303 yenGross loss: -32 yenPayoff ratio: 2.37Max reported daily loss: -32 yen ■ LLMBridgeTrader -816 yenRecord: 1 positive day / 4 negative daysDay-level win rate: 20.0%Gross profit: +33 yenGross loss: -849 yenPayoff ratio: 0.16Max reported daily loss: -466 yen ■ MLScore GF-T4 -1,578 yenRecord: 0 positive days / 4 negative days / 1 flat dayDay-level win rate: 0.0%Gross profit: +0 yenGross loss: -1,578 yenPayoff ratio: N/AMax reported single-trade loss: -602 yen ■ Total -3,490 yenRecord: 7 positive bot-days / 12 negative bot-days / 1 flat bot-dayDay-level win rate: 36.8% excluding flat resultGross profit: +539 yenGross loss: -4,029 yenPayoff ratio: 0.23Max reported daily loss: -712 yen Today’s theme: exits beat win rate The main theme this time is not portfolio diversification. It is exit quality. A bot can filter entries, avoid bad setups, and still lose if the exit logic lets one bad position grow beyond the size of many normal wins. GateGrid AI is the clearest example. Its design is built around not entering weak conditions: CatBoost checks the gate first, then Ollama can return defensive decisions such as AI_SKIP(sess=NY gate=0.50 base_thr=0.54 adj_thr=0.55) or OLLAMA_HOLD. That kind of log is useful because it tells me the machine is not blindly firing orders. But the five-day result says the harder problem sits after the entry: once a position or basket survives the filters, the loss needs to be cut before it becomes the whole story. Bot design notes describe GateGrid AI as a CatBoost + Ollama multi-gate system, while BoundSniper mainly relays TradingView signals to MT5 and LLMBridgeTrader asks AI to output OPEN/HOLD/CLOSE/REVERSE style position actions. GateGrid AI GateGrid AI ended at -1,367 yen. The raw daily path was +67, -703, -712, -155, +136 yen. The last day recovered a little, but the middle of the week had already done the damage. The frustrating part is that the bot is not reckless by design. It is supposed to block weak entries with CatBoost and then ask Ollama to judge the environment with spread, ATR, higher-timeframe trend, session, recent win rate, and recent P/L. That is a good structure on paper. Still, the result looked like a classic small-win, large-loss pattern. The entry gate may be doing something useful, but the basket exit is probably still too forgiving. I do not have full certainty yet, but that is where my eyes go first. A payoff ratio of 0.19 on the day-level summary is a warning sign. I know this is not the exact trade-level payoff ratio, but the shape is hard to ignore. If the average losing day is five times the average winning day, a high internal win rate becomes less comforting very quickly. BoundSniper BoundSniper finished at +271 yen, and it did it with the least dramatic architecture. This bot is basically an execution bridge for TradingView signals on USDJPY. It does not try to be clever about the market itself. That simplicity helped. The daily path was +182, +3, -32, +20, +98 yen. No huge win, no heroic AI judgment, no long explanation needed. The largest negative day was only -32 yen, which almost feels boring, but boring was valuable this week. The payoff ratio came out at 2.37 on a day-result basis. That is the only bot where the loss side did not dominate the week. I would not overpraise it from five days of data, but in this window it behaved like the adult in the room. LLMBridgeTrader LLMBridgeTrader ended at -816 yen. The daily line was +33, around -261, -466, -92, -30 yen. Not pretty, but the loss profile is different from GateGrid AI. This bot matters because it asks AI to manage more than direction. It can choose OPEN, HOLD, CLOSE, REVERSE, or NONE, and it also produces confidence, setup type, SL pips, TP pips, entry reason, and exit reason. In theory, that gives it a better chance to escape bad positions by switching from holding to closing. In this five-day block, the result still landed negative, but the last two days were relatively contained at -92 and -30 yen. That does not prove the exit logic works, though it hints that the damage may be more controlled than the headline win rate suggests. I would keep watching the CLOSE and REVERSE decisions. If this bot is going to become useful, the edge will probably come less from calling direction perfectly and more from admitting the trade is no longer worth holding. MLScore GF-T4 MLScore GF-T4 was the heaviest drag after GateGrid AI, finishing at -1,578 yen. The reported path was 0, -487, -405, about -234, and -452 yen. There was no positive day in the period. The entry-blocking function seems to be doing part of its job, but the trades that do get through carry too much downside. The note that June 19 had one losing trade around -602 yen is the kind of number that changes the feel of the whole bot. I saw that and thought, not again with the wide stop. This bot may not need more intelligence first. It may need a smaller permission space. Fewer trades are not enough if the approved trades can still hit a loss size that the rest of the portfolio cannot absorb. Summary The week ended negative, but the useful finding is clear: the best-looking AI structure did not automatically create the best risk structure. GateGrid AI had smart filters and still lost badly. LLMBridgeTrader had richer decision language and still could not climb out. MLScore GF-T4 blocked some entries but let too much loss through when it acted. BoundSniper, the simple rule-based execution bot, was the only one that left the week positive. I do not think that means “AI failed” in some grand way. It means the experiment has moved from entry quality to exit discipline. The next improvement should probably be less about asking the model to be smarter, and more about making it harder for any one trade or basket to become the whole week. 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

2 min.
20 jun

A 72.2% Win Rate Wasn’t Enough: One Oversized Loss Changed the Whole Day

A 72.2% win rate should have made this a decent trading day. It did not. Across four MT5 bots, the combined record was 13 wins and 5 losses, but the final result was -248 yen. For this log, I am treating June 19, 2026 as Day 1 of the published series. Cumulative P/L is now -248 yen. The important part was not that the bots failed to find winners. They found plenty. The problem was that the average loss was too large compared with the average win. Bot results ■ GateGrid AI +136 yenSymbol: GBPUSD-Record: 7W / 1LWin rate: 87.5%Gross profit: +296 yenGross loss: -160 yenPayoff ratio: 0.26Max loss: -160 yen ■ BoundSniper +98 yenSymbol: USDJPY-Record: 3W / 0LWin rate: 100.0%Gross profit: +98 yenGross loss: 0 yenPayoff ratio: N/AMax loss: 0 yen ■ LLMBridgeTrader -30 yenSymbol: EURUSD-Record: 2W / 3LWin rate: 40.0%Gross profit: +254 yenGross loss: -284 yenPayoff ratio: 1.34Max loss: -132 yen ■ MLScore GF-T4 GBPJPY -452 yenSymbol: GBPJPY-Record: 1W / 1LWin rate: 50.0%Gross profit: +150 yenGross loss: -602 yenPayoff ratio: 0.25Max loss: -602 yen ■ Total -248 yenRecord: 13W / 5LWin rate: 72.2%Gross profit: +798 yenGross loss: -1,046 yenPayoff ratio: 0.29Max loss: -602 yenRunning day: Day 1Cumulative P/L: -248 yen Today’s theme: the win rate looked fine, the payoff ratio did not The headline number was 72.2%. On paper, that sounds like a day where the bots were mostly right. But the average winning trade was about 61 yen, while the average losing trade was about 209 yen. That is the part that explains the result. The bots did not need many losing trades to finish negative. One -602 yen loss from MLScore GF-T4 GBPJPY was enough to offset a large part of the smaller wins from the other systems. This was a day where the hit rate gave a comfortable impression, but the payoff structure told a different story. GateGrid AI: profitable, but still dependent on small wins holding up GateGrid AI finished at +136 yen, the best net result among the four bots. The record was 7 wins and 1 loss, with an 87.5% win rate. That looks strong, but the payoff ratio was only 0.26, so the wins were frequent and small. The AI layer was active. One log line showed the system allowing a BUY grid with sig=BUY, conf=0.75, timing=TREND_FOLLOW, and a reason beginning with “ML DEFENSIVE.” That tells me the bot was not simply stacking orders without a filter. It was passing through a mix of ML scoring and LLM judgment. The exit side still needs attention. The log repeatedly showed: “price hit but pnl=-348.00 BoundSniper: simple execution, clean result BoundSniper finished at +98 yen with 3 wins and no losses. This bot is not trying to interpret the market through an LLM. Its job is to receive TradingView signals through a webhook and execute them on MT5. The live monitor showed the webhook server, cloudflared tunnel, and MT5 trader all running, with public health marked OK. That is not a dramatic trading insight, but it matters. A signal relay bot first has to be reliable as infrastructure. The profit was small, but the result was clean. BoundSniper did not add unnecessary complexity to the day. It received signals, executed them, and closed positive. LLMBridgeTrader: lower win rate, better loss shape LLMBridgeTrader finished at -30 yen. The win rate was only 40.0%, with 2 wins and 3 losses, but the payoff ratio was 1.34. That makes this bot more interesting than the final P/L suggests. The AI log included the exit reason: “Close managed position as strong downtrend invalidates reversal setup.” This is the part worth following over time. The bot is not only choosing entries; it is also explaining when the original trade idea no longer holds. There were also mean-reversion BUY ideas inside a strong bearish environment. One log phrase was: “Oversold RSI 16 in strong bear trend indicates potential pullback.” That logic is understandable, but it can be fragile. Oversold conditions inside a trend do not always mean reversal. Still, the largest loss was -132 yen, so the damage was contained compared with the GBPJPY bot. MLScore GF-T4 GBPJPY: the one trade that changed the portfolio MLScore GF-T4 GBPJPY was the main drag on the day. It had only two closed trades, one win and one loss. The win was +150 yen, while the loss was -602 yen. That left the bot at -452 yen. The score log showed candidate=SELL decision=ENTER signal=SELL score=93.5 trade=done. The score was high, the trade was accepted, and the system entered. The result later became the largest loss of the day. This does not mean the score model is broken. A high-score setup can still lose. The issue is that the loss size was too large relative to the win size. For GBPJPY, the next review should probably focus less on signal confidence and more on stop size, volatility adjustment, and daily loss limits. Summary Day 1 ended at -248 yen. Two bots finished positive, one was only slightly negative, and one large GBPJPY loss pulled the total below zero. The portfolio did not lose because it could not win trades. It lost because the losing trades were too heavy. The result is a useful starting point for the series. Entry accuracy is visible and easy to talk about, but the exit design is where the real test begins. 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

2 min.
18 jun

Win Rate Lied Today: 4 MT5 LLM Bots Finished at -464 Yen Despite a 75% Win Rate

The result for June 18 was simple, but a little uncomfortable: four MT5 auto-trading bots closed the day at -464 yen. The portfolio won 27 out of 36 trades, so the surface number looked fine. Still, the payoff ratio was only 0.27, and that is where the day fell apart. This was not a day where the bots failed to find entries. They found plenty. The real question was whether they knew when to stop believing those entries. I think the problem sits mostly at the exit, though the evidence is not perfectly clean yet. Bot-by-bot results ■ GateGrid AI -155 yenSymbol: GBPUSD-Closed trades: 21Record: 17W / 4LWin rate: 81.0%Gross profit: +955 yenGross loss: -1,110 yenPayoff ratio: 0.20Max loss: -443 yen ■ BoundSniper Bot +20 yenSymbol: USDJPY-Closed trades: 7Record: 6W / 1LWin rate: 85.7%Gross profit: +446 yenGross loss: -426 yenPayoff ratio: 0.17Max loss: -426 yen ■ LLMBridgeTrader -92 yenSymbol: EURUSD-Closed trades: 5Record: 3W / 2LWin rate: 60.0%Gross profit: +348 yenGross loss: -440 yenPayoff ratio: 0.53Max loss: -241 yen ■ MLScore GF-T4 GB -237 yenSymbol: GBPJPY-Closed trades: 3Record: 1W / 2LWin rate: 33.3%Gross profit: +265 yenGross loss: -502 yenPayoff ratio: 1.06Max loss: -252 yen ■ Total -464 yenClosed trades: 36Record: 27W / 9LWin rate: 75.0%Gross profit: +2,014 yenGross loss: -2,478 yenPayoff ratio: 0.27Max loss: -443 yenSeries day: Day 1 in this provided logCumulative P&L: -464 yen Today’s theme: entry confidence did not protect the downside The most interesting contrast was not between winners and losers. It was between win rate and damage size. GateGrid AI won more than 80% of its exits, but still ended negative. BoundSniper Bot won almost everything and barely escaped positive. Seeing a -443 yen single loss inside GateGrid made me pause, because the rest of the day was mostly small repairs. The ML score bot was even more direct. The GBPJPY engine printed candidate=SELL decision=ENTER signal=SELL score=95.09 trade=done, but that first SELL later closed at -252 yen. A score over 95 feels strong on screen. In the account report, it was just the first of two stop losses before the final win. GateGrid AI GateGrid AI’s final number was -155 yen, but that result hides a rough start. The first two closed trades were -379 yen and -443 yen, and the log says it plainly: [CAP] Basket max-loss exit: pnl=-822.00 The win rate was 81.0%, which sounds good until the payoff ratio shows up at 0.20. Average winners were small; average losers were too heavy. This is the kind of day where the grid logic can look active and controlled, while the account curve is still paying for one bad pocket of exposure. BoundSniper Bot BoundSniper Bot was the only bot to finish positive, but only by +20 yen. That number is almost funny after a 6W / 1L record. I had to look twice, because a win rate of 85.7% usually feels like it should leave more behind. The explanation is blunt: six wins made +446 yen, and one loss took back -426 yen. Payoff ratio was 0.17. BoundSniper is not using the same kind of LLM planning as LLMBridgeTrader, so I do not want to force the “AI exit” story onto it. But as a trading system, it is showing the same design issue: one bad exit can consume a clean run of small closes. LLMBridgeTrader LLMBridgeTrader lost -92 yen, but it gave the most useful AI log of the day. Near the final EURUSD trade, the model opened a BUY with a mean-reversion setup: RSI oversold bounce in weak ranging market. Then, fifteen minutes later, it chose to close with position_action":"CLOSE" and the exit reason Weak ADX and bearish structure favor closing. That close finished as a +90 yen trade. This is the part I like. The bot was not only saying BUY or SELL. It was also changing its mind about holding the position. The full day was still negative because the earlier two losses were larger than the later wins, but the exit behavior itself was not random. It saw weak trend conditions and got out. That is closer to the experiment I want this bot to run. MLScore GF-T4 GB MLScore GF-T4 GB had the lowest win rate at 33.3%, yet its payoff ratio was the best of the group at 1.06. That sounds contradictory, but it fits the trades: two losses around -250 yen, then one win at +265 yen. The structure was not terrible. The timing was. The entries were all SELL signals with strong scores, including 95.09, 92.89, and 93.5. The issue is that high score did not mean high follow-through. If this bot keeps using score as the main gate, the next layer probably has to judge when a high-score setup is already late. Wrap-up The day ended at -464 yen. Not a disaster, but not a clean loss either. The strange part is that the portfolio was right often enough. It just did not get paid enough when right, and it paid too much when wrong. For this series, today’s lesson is not “raise the win rate.” The bots already did that. The better question is whether each bot can recognize the moment when its original idea has expired. That is where the next improvement probably lives. 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

14 min.

Bekijk alles (38)

Can AI really trade forex? AI FX Bot Lab is a real-time experiment from Japan, where I build and test AI-assisted FX trading bots using MT5, Python, machine learning, and local LLM tools. I share live results, failures, risk lessons, and bot improvements from rule-based, AI-driven, and ML + LLM hybrid systems. Not financial advice. fxaibotlab.substack.com

Maker

Kimi | Japan FX Bot Lab
Jaren actief

2K
Afleveringen

38
Beoordeling

Veilig
Copyright

© Kimi | Japan FX Bot Lab
Website van aflevering

AI FX Bot Lab: Real Trading Experiments