SACHI_USDJPY_M5_GPU（GPU対応の適応学習型EA）

SACHI_USDJPY_M5_GPU is an MT5 Expert Advisor designed for short-term trading on the USDJPY 5-minute chart.
This EA does not trade merely on simple indicator crosses or fixed conditions, but evaluates price movement, volatility, candlestick positions, short-term trends, range conditions, volume-equivalent information, recent win/loss tendency, spreads, and market state to determine whether to buy, sell, or stay out.

The development concept is not aiming for flashy one-shot profits, but to narrow trading according to market conditions, avoid forced entries, and reduce losses while increasing profits.
Especially on USDJPY M5, there are many situations where direction changes quickly, moves within a range with up and down fluctuations, and short-term volatility spikes that temporarily disrupt spreads and volatility. To reduce such short-term trading failures, this EA combines multiple guard features and market determination logic.

This EA supports extensions such as Local AI, GPU, and OpenBLAS-based L1 trend filters, but these are not mandatory.
Even in environments without AI or GPU, the core EA trading logic is designed to operate. Local AI does not completely replace trading decisions; it serves as auxiliary judgment, risk adjustment, additional chart evaluation, and monitoring of loss patterns.

This EA is adjusted for use on USDJPY M5.
Parameters can AUTO specify the symbol, but the logic, thresholds, volatility determination, news/intervention warnings, spread checks, and market regime determinations are mainly built around USDJPY 5-minute charts. Therefore, the recommended environment is USDJPY M5 chart.

Applying to other currency pairs or other timeframes is sometimes possible, but expected behavior and risk characteristics will change, so the sales page only states USDJPY M5-specific or USDJPY M5 recommended usage.

This EA classifies the market into major categories such as trend, range, high volatility, and unclear conditions, and selects the most suitable candidate at that time.
Entry candidates include pullbacks, mean reversion, short-term reversals, breakout-leaning candidates, and conservative candidates, with their priority adjusted according to the current market state.

For example, in a strong trend, it avoids simple counter-trend entries and prioritizes pullbacks and retracements in line with the flow.
Meanwhile, in a range with weak direction, it avoids aggressive breakout targets and excessively extended take-profit goals outside the range.
Also, after sudden moves or when spreads widen, judgments are made more cautiously than usual.

In entry decision, characteristic features for both buy and sell directions are calculated, and a comprehensive judgment is made based on probability, candidate score, market regime, recent profit/loss trend, AI adjustment, spread, volatility, and news/market state.v9.10 adds multi-layer entry assistance that checks M15 volume, M15 EMA20/60, H4 EMA50, ADX/RSI, and the last 14 weighted scores, in addition to the conventional candidate selection. This judgment is not a hard filter that mechanically stops trades but a soft filter that favorably evaluates candidates that meet conditions and cautiously adjusts insufficient candidates.

It is not built to enter trades based on a single condition; orders are only placed after passing multiple evaluations.

Design for Profit Preservation and Loss Limitation

This EA does not place stop-loss and take-profit solely based on fixed pips, but uses dynamic SL/TP based on ATR.

v9.10 adds Dynamic RR where, when H1 RSI, AI direction bias, market bias, L1 trend, and sufficient candidate probability are strong, the initial TP ATR multiplier is automatically increased. In normal conditions it maintains the traditional ATR-based TP, expanding only in situations with high potential for extension.

When market volatility is high, price ranges widen take-profit, and when volatility is low, it reduces take-profit, aligning the profit target with the market environment.

In the initial setup, stop-loss and take-profit are calculated by ATR multipliers.
Furthermore, during position holding, it aims to manage profits favorably by combining break-even, trailing stop, partial profit-taking, and additional profit extension processes.

In short-term trading, taking profits too early can cause missed upside, while waiting too long can erode profits. This EA manages risk gradually by considering volatility, current price, strength of candidates, range width, and current holding status rather than fixing a large range from the start.

Handling of range markets

In USDJPY short-term charts, there are many periods of flat movement within a certain range.
In such cases, setting take-profit that far exceeds the range may lead to reversal before reaching it and loss of profits.

This EA checks recent highs/lows, ATR, price position, and range progression, and when it detects a range market, it avoids excessive take-profit targets.
It does not set abnormally distant TP relative to the range width, aiming to reduce “unreachable take-profits” in flat markets.

Thus, it shifts to extending profits in trending markets while avoiding overextension in range markets.

Anti-surge and intervention warning logic for yen appreciation

This EA includes a yen intervention warning logic to detect USDJPY-specific rapid declines.
When the yen strengthens sharply within a short period, price movements can diverge from typical technical markets. A simple pullback or contrarian entry may risk being swept up in a sudden drop.

The EA detects sharp declines on the M5 chart through the extent of the drop, multiple consecutive down moves, changes in volume-equivalent, spread, and comparison with ATR.
If detected, it suppresses risk on existing long positions and temporarily controls regular entry judgments.

Also, after a drop, it does not merely stop trading; if there are signs of rebound near the bottom, it can evaluate buy-back candidates as well.
However, this is not a rule to buy on sharp declines unconditionally. It checks decline magnitude, rebound from the low, spread, and distance from the low to avoid jumping in during extreme drops.

Loss-streak suppression and re-entry logic

This EA includes guards to suppress forced re-entries during consecutive losses.
In short-term EAs, trading during times when market conditions do not fit the logic tends to enlarge losses. Therefore, after a certain number of consecutive losses, it switches to more cautious judgments.

On the other hand, if the loss-streak guard is too strong, opportunities after market changes may be missed.
Thus, this EA not only resets with date changes but also has a mechanism to release guards when chart conditions change.
It checks price range, volatility, recent price movement, and changes from highs and lows, and if it can determine that the market has shifted away from a loss-streak phase, it re-evaluates entry candidates again.

AI auxiliary features

This EA has functionality to connect with a local AI server.
The modeled assumption uses lightweight VLM models such as qwen2.5-vl-3b-instruct, and can utilize OpenAI-compatible local servers like LM Studio.

The role of AI is not to completely replace the EA’s own trading decisions.
The core remains the MQL5-based trading logic, and AI provides auxiliary judgments such as:

brief market explanations, risk adjustment, buy/sell directional bias, auxiliary evaluation of breakout versus pullback/reversal, decision to adopt or observe current candidates, monitoring loss patterns of held positions, etc.

Even if AI cannot connect, the EA’s core logic will operate.
When AI is connected, the panel displays AI status, AI reason, AI confidence, AI bias, and AI adopted, showing what auxiliary judgments AI is currently providing.

Loss pattern monitoring using chart images

The EA uses local AI’s image recognition to monitor the chart state of held positions.
If a losing trade occurs, the chart image and state at that time can be recorded and referenced as a loss pattern by the local AI.

During position holding, charts are periodically sent to the local AI to check if the current shape resembles past loss patterns.
If it is highly likely to be a loss pattern and the position is profitable, it can decide to secure profits.

This feature aims to assist in detecting “apparent breakdowns” or “past similar losses” that are difficult to capture with only numeric logic.
However, AI judgments are not absolute, so the EA’s core checks confidence and profit conditions to avoid overreaction.

Saving VLM learning data

The EA includes a feature to save information at the time an entry candidate is confirmed as VLM learning data.
Objects saved include symbol, direction, candidate ID, probability, regime, buying-direction features, selling-direction features, etc.

This allows later local AI or external analysis environments to review under what conditions the EA selected entry candidates.
It can be used in the future to enhance AI decision accuracy or to validate using chart images and numerical features together.

GPU / OpenCL support

This EA supports GPU usage via OpenCL.
In environments where GPU is available, feature evaluation and model calculations can be accelerated to reduce internal EA processing load.

It is designed to run without GPU as well, but for comfortable use of local AI and image recognition, a GPU such as NVIDIA RTX series is recommended.
For the EA’s standard logic alone, it can run on a typical MT5-compatible VPS.

OpenBLAS L1 trend filter support

The EA includes trend-detection using MQL5’s OpenBLAS-based L1 trend filter.
L1 trend filter not only smooths price like a simple moving average but also makes it easier to extract the market slope and turning points more naturally.

This feature allows the EA to evaluate current trend direction, slope strength, and turning signals while reducing overreaction to short-term noise.

Within the EA, results are reflected in candidate scores and features to suppress aggressive entries against the trend and to evaluate reversal candidates.

Risk management features

The EA includes multiple risk management features.
Main features include ATR-based SL/TP, break-even, trailing stop, partial profit-taking, maximum losing-streak guard, daily P/L and daily drawdown display, spread block, market close consideration, trading restrictions before/after news, and restrictions on additional entries during position holding.

The initial setting operates with a single position; it is not designed on the premise of excessive averaging down or Martingale.
Fixed lot operation is supported, but it is recommended to start with small lots according to account funds and strategy, and to verify with a demo account or strategy tester before live use.

This EA does not eliminate risk in short-term trading.
Market sudden changes, spread widening, order execution delays, server environment, local AI response delays, and communication conditions can affect actual results.

Panel display

An information panel displaying the EA’s status is shown on the right side of the chart.
The panel shows the current top blockers, operational status, GPU status, probability, candidate selection results, entry gate, learning status, final action, signals, number of candidates passed, selection probability, account balance, daily P/L, daily drawdown, spread, ATR, volume, consecutive losses, AI status, and AI reasons.

The panel consists of multiple pages, where you can also view detailed candidate evaluations, AI judgments, vision tests, and timestamp information.
Since it visualizes why the EA did not enter, which guard stopped it, and how AI is judging, it makes monitoring the running state easier. Also, due to the large amount of information,InpTradeExecutionSuppressMode（Trade execution suppress mode） set to truewill suppress automatic trading and allow use as an indicator for discretionary trading.

Recommended operating environment

When using only the standard logic, MT5 can operate on a stable Windows environment.
For full feature utilization, Windows 10/11 64-bit, latest MetaTrader 5, CPU with 6 cores or more, 32GB RAM or more, 20GB SSD free space or more, NVIDIA RTX 3060 12GB or more is recommended.
If you also use the local AI image recognition function, a GPU such as RTX 3090 24GB or greater is even more comfortable.

If using local AI, start a local server compatible with OpenAI such as LM Studio and use a VLM model like qwen2.5-vl-3b-instruct.
Also, you need to register the local AI server URL in the WebRequest allowed URLs in MT5 options.

GPU and local AI are not mandatory.
Even if these are disabled, the EA’s core logic will still operate.

※ Before using this EA, please read and agree to the disclaimers described in the manual. If you cannot agree, please refrain from use. It is strongly recommended to understand investment risks when using this EA.