AIBridge USDJPY

Product Overview Section

What the EA can do

AIBridge USDJPY is anautomated trading programthat runs on. It learns from past chart patterns using artificial intelligence (AI) to judge and trade the market 24 hours a day.MetaTrader 4

• 27-dimensional technical features（EMA・RSI・MACD・ATR・ADX・Bollinger Bands, etc.）
• Time zones and session information（Tokyo・London・NY・overlap times）
• Automatic market regime detection（Trend / Range）
• US 10-year yield (FRED) integration（Highest correlation indicator for USDJPY）
• Machine learning inference with LightGBM（Backtest and live operation use the same model）

What makes it different from ordinary EAs?

Ordinary EA:

“Buy when the moving average trends up”follows simple rules decided by humans. Because the rules are fixed, when market behavior changes, it stops working quickly.

AIBridge:

“Analyze over 6.5 years of charts (more than 40,000 candlesticks) and U.S. interest rate dataAI”,analyzes and learns the pattern,”such as “in this situation, prices tended to move this way in the past,”. It learns on its own. Instead of a human deciding “when to buy,”AI makes the optimal decision according to the situation.

🤖 What does AI (machine learning) do?

When you hear “AI” or “machine learning,” it sounds hard, but what it does isthe same as school study.

If you had to compare it, it’s like “an outstanding student who has solved a lot of past papers”

How a human student solves past problems:
1. Solve problems from the past 10 years
2. Memorize “when this kind of problem appears, solve it like this”
3. When a similar problem appears on the actual test, you can get it right
AI learning flow:
1. Show about 40,000 candlesticks from the last 6.5 years
2. Learn how prices moved after certain movements plus interest rate conditions
3. In production (live operation), when a similar pattern occurs, determine using past data

Hints AI is looking at (features)

AI, at any chart point, looks at about 40 different numerical values simultaneously: Content

Category	Content
📈 Price movement	Recent 1–20 candles up or down rates
🌊 Trend indicators	Moving averages, MACD, EMA crosses
📊 Momentum indicators	RSI, momentum
⚡ Volatility	ATR, Bollinger Band width
🕐 Time of day	Activity times of Tokyo/London/NY markets
📅 Day of week	Cyclical patterns like Mondays volatile, Friday afternoons quiet
🇺🇸 U.S. interest rate level	US 10-year yield（External factor most affecting USDJPY）

Ordinary EA looks at only 1–3 indicators, but AIBridgejudges 40 indicators simultaneously. This is a feat humans cannot do and is the strength of AI.

What to be wary of when you hear AI is learning

⚠️AI only remembers past patterns; it does not predict the future.

When a pattern never seen before (like Lehman-style crash, war, major policy shifts) occurs, AI may not respond.Do not overtrust— understand it as a probabilistic tool based on past tendencies.

---

Differences from conventional EAs

Aspect	Typical EA	AIBridge USDJPY
Logic	Fixed rules (moving averages, etc.)	Dynamic judgment by AI
Number of inputs	1–3	Refer to about 40 at once
US interest rate info	Not considered	Always referenced
Parameter tuning	Manual by user	Built-in trained AI
Backtest method	In-sample (risk of overfitting)	Walk-Forward validated
Transparency	Numbers only	All trade history公開

Performance (actual results)

Walk-Forward Analysis – validating without overfitting

🔬 What is Walk-Forward Analysis?

“Backtesting method that eliminates overfitting and is the most reliable”

Problems with ordinary backtests

Ordinary backtests are like “taking a test while peeking at the answers”:“Past 6 years’ answers are all known to AI”. Then you test on the same six years and:

Normal backtest (In-sample):
① Teach AI all the answers from the past 6 years
② Test on the same 6 years
③ “Amazing! 100% correct!”
→ This does not tell you if it will be right in the future

In reality, ordinary backtests can yield annualized returns over 50%, but live operation almost certainly ends in a big failure.

AIBridge was validated with the “correct method”

Walk-Forward Analysis:
Step 1: Train AI with data from 2019/10–2023/10
Step 2: Test with data 2023/10–2024/01 (future data not used for training)
Step 3: Advance training data to 2020/01–2024/01
Step 4: Test with 2024/01–2024/04 (future data again)
Step 5: (repeat) 
→ All performance measured using only data not used for training
→ Real-world conditions = reliable performance

Illustration

[Training] 2019/10 ───── 2023/10 [Test] 2023/10 ─ 2024/01
[Training] 2020/01 ───── 2024/01 [Test] 2024/01 ─ 2024/04
[Training] 2020/04 ───── 2024/04 [Test] 2024/04 ─ 2024/07
... (9 cycles)

AIBridge Walk-Forward results

Testing period: October 2023 – January 2026 (27 months)
Number of tests: 9 (slid every 3 months)
Aggregate net: +36.73%
Profit period: 8/9 (89%) profitable
Best period: +10.21% (Jan–Apr 2025)
Worst period: -0.98% (Jul–Oct 2025)
Win rate: 48.5%
Profit factor: 1.77
Sharpe ratio: average 2.5
Max DD: under 5% on a folding basis

Asset growth chart

Walk-Forward Analysis (training data up to Oct 2023) (Out-of-Sample)
※ Trade for about two years into the future all at once

Backtest after full training (In-Sample)

Monthly profit/loss by Walk-Forward analysisShare of monthly P/L

A test of 1000 scenarios for “what if this system had been operated again under the same conditions?”

Questions Monte Carlo answers

“Was it just lucky to get good results this time?” “How big could the loss be in the worst case?” “What happens if a drawdown continues during bad luck?”

Monte Carlo is the statistical method that answers these questions.

Idea: simulate 1000 possible worlds

Trades executed in the live backtest:
[+200, -100, +150, -80, +300, -50, ...] ← just a lucky dataset
Monte Carlo tests 1000 “what if” scenarios:
What-if world 1: [+150, +200, -50, -100, +300, -80, ...] → Final +1,000
What-if world 2: [-100, +200, +300, -50, +150, -80, ...] → Final +800
What-if world 3: [-80, -100, -50, +300, +200, +150, ...] → Final -200  ← unlucky pattern
...
What-if world 1000: ...
→ From 1000 patterns, compute statistically the typical range and the best/worst cases

Key numbers from Monte Carlo

AIBridge Monte Carlo results (1000 runs)

Final capital (start from $10,000):
5% Worst case:      $624,218  ← worst-case with bad luck
50% Median:           $697,008  ← typical result
95% Best case:       $771,094  ← favorable result
Probability of ending in loss: 0.00%   ← no capital loss risk
Maximum drawdown:
5% (favorable path):  7.48%
50% (typical):         15.68%
95% (unfavorable path): 38.71%   ← expect this level of DD
99% (extremely rare):     59.47%   ← DD beyond this is <1%

Monte Carlo distribution

Walk-Forward Analysis（training data up to Oct 2023; Out-of-Sample）

※ Trade for about two years into the future all at once

Operating Environment

Required environment

• MetaTrader 4（Broker optional, supports demo accounts）
• Windows VPS（recommended） or Windows PC
• Python 3.10 or higher（installer included）
• Internet connection（for US 10-year yield data, once per day）

Recommended specs

Recommended broker conditions

Set contents

✅ AIBridgeEA.ex4 (compiled EA file) ✅ AIBridgeEA.mq4 (EA source code) ✅ Python server (installer) ✅ Trained AI model (27-dimensional features + US 10-year yield integration) ✅ Historical update tool (automatic US 10-year yield data retrieval) ✅ Detailed setup manual (PDF) ✅ Backtest and Walk-Forward Analysis results bundle (CSV + graphs)

FAQ

Q. Why can’t this be used with EUR/USD or other currencies?

A. AIBridge is optimized specifically for USDJPY. It uses the high correlation between USDJPY and US 10-year yield in learning, so it does not achieve the same performance with other currencies.Prioritizes specialized performance over versatility.

Q. Do I need knowledge of Python?

Q. Are there specific brokers?

A. It runs with any broker that supports MetaTrader 4. HoweverECN/STP accounts with spreads under 1.5 pips are recommended. Wider spreads significantly reduce performance.

Q. How should I handle money management?

A.Default settings: risk 1% per trade, up to 1 position, maximum daily drawdown 3%. If capital is small, reduce to 0.5% and increase back to 1% as you gain experience. Money management parameters are adjustable.

Q. Is MT5 supported?

A.Currently only MT4 is supported. MT5 version is under consideration for the future.

Q. What about news events?

A. The system includesnews suppression logic that automatically suppresses opposite signals during strong news releases. However, it cannot be completely prevented. It is also recommended to pause the EA manually before and after major news releases.

Q. What if the EA stops due to an error?

A.Existing positions are safely managed on MT4 (trailing, breakeven, and stop losses are handled on MT4). Rebooting will automatically recover.