Hello traders and fellow enthusiasts,
We’re happy to join your forum and share our journey in building a Forex prediction system powered by AI. In this post, we’ll walk you through how we increased our forecast accuracy from 50% to 80%, step by step — and what we plan to do next.
Quick note:
• Training data is the historical market data used to train the AI.
• Test data is unseen data used to evaluate prediction performance (as close to real market conditions as possible).
⸻
✅ Stage 1: Finding the Most Effective Input Data (+12% Accuracy)
We started by experimenting with the types of inputs we feed into the neural network:
• Number of Japanese candlesticks (bars)
• How far ahead we try to predict
• Timeframes (M15, H1, H4, etc.)
After testing multiple configurations, we discovered the best results came from:
• 2500 input candlesticks
• Forecasting 50 bars ahead
• 1-hour timeframe
Accuracy: 62%
Interestingly, using fewer candlesticks reduced accuracy, while adding more didn’t help — and slowed the system down. So we optimized for efficiency.
⸻
Stage 2: Designing Better Neural Network Architectures (+9% Accuracy)
Initially, our neural network performed exceptionally well on the training data — achieving up to 99% accuracy — but only 62% on the test data, which revealed a major issue with overfitting.
To address this, we began testing various neural network architectures, including known designs and our own custom ideas.
Through trial and error, we found that our custom architecture delivered better results and generalized more effectively.
We ran the training process over 250+ separate restarts, each time tweaking architecture parameters, retraining from scratch, and analyzing performance metrics.
By comparing all results, we identified one architecture that consistently outperformed the rest — and later refined it even further for efficiency and speed.
This led to an accuracy improvement on real (unseen) data from 62% to 71%.
⸻
Stage 3: Identifying the Best Currency Pairs (+3% Accuracy)
(This stage both helped and hurt us — you’ll see why in a next stage)
At first, we aimed to create a universal neural network trained on all currency pairs, so it could provide predictions across the entire market.
However, during training, we noticed a pattern:
Neural networks trained on just a single currency pair consistently outperformed those trained on multiple pairs.
After analyzing the results, we discovered that:
• Each currency pair has its own unique behavioral patterns
• While some general patterns exist, cross-pair noise reduced accuracy
So we shifted our focus — training separate models for each currency pair.
This boosted accuracy from 71% to 74%.
But here’s the downside:
To support multiple currency pairs now, we must train and maintain a dedicated neural network for each pair, which increases system complexity and resource usage.
So yes — this stage gave us a nice gain in accuracy, but also introduced new challenges.
⸻
Stage 4: Merging Models for a Synergistic Effect (+3% Accuracy)
Here’s the twist:
By chance, we tested a combination of two models:
• One trained on a specific pair
• Another trained on all pairs
We observed that:
• Their combined predictions were less frequent but more accurate.
• The diversity in pattern recognition helped the model filter out noise.
Resulting accuracy: 77%
⸻
️ Stage 5: Data Augmentation, Dropout & Noise (+3% Accuracy)
As the model kept training, we noticed a problem — it began memorizing the data, and accuracy on test sets dropped.
So we implemented:
1. Dropout — randomly disables some neurons during training to prevent overfitting.
2. Input scaling — adds small random variations to inputs, encouraging pattern recognition over memorization.
3. Noise injection — we add slight random noise to candlestick data, training the AI to handle imperfect inputs.
Final accuracy reached: 80% on unseen, real-world test sets.
⸻
What’s Next?
We know that candlestick patterns alone aren’t enough.
The market is influenced by:
• Economic news
• Political events
• Unexpected external shocks
Our next step?
We’re building an AI that will read and analyze financial news in real time, and combine this with our current market pattern model — aiming for even better predictive power.
⸻
Thanks for reading!
We’d love to hear your thoughts, feedback, or suggestions.
Have you tried similar approaches? What worked for you — and what didn’t?
If you’re interested in testing the app, here’s the link:
https://play.google.com/store/apps/details?id=com.PersianDare.AITrading
Let’s build the future of trading together.
We’re happy to join your forum and share our journey in building a Forex prediction system powered by AI. In this post, we’ll walk you through how we increased our forecast accuracy from 50% to 80%, step by step — and what we plan to do next.
Quick note:
• Training data is the historical market data used to train the AI.
• Test data is unseen data used to evaluate prediction performance (as close to real market conditions as possible).
⸻
✅ Stage 1: Finding the Most Effective Input Data (+12% Accuracy)
We started by experimenting with the types of inputs we feed into the neural network:
• Number of Japanese candlesticks (bars)
• How far ahead we try to predict
• Timeframes (M15, H1, H4, etc.)
After testing multiple configurations, we discovered the best results came from:
• 2500 input candlesticks
• Forecasting 50 bars ahead
• 1-hour timeframe
Accuracy: 62%
Interestingly, using fewer candlesticks reduced accuracy, while adding more didn’t help — and slowed the system down. So we optimized for efficiency.
⸻
Stage 2: Designing Better Neural Network Architectures (+9% Accuracy)
Initially, our neural network performed exceptionally well on the training data — achieving up to 99% accuracy — but only 62% on the test data, which revealed a major issue with overfitting.
To address this, we began testing various neural network architectures, including known designs and our own custom ideas.
Through trial and error, we found that our custom architecture delivered better results and generalized more effectively.
We ran the training process over 250+ separate restarts, each time tweaking architecture parameters, retraining from scratch, and analyzing performance metrics.
By comparing all results, we identified one architecture that consistently outperformed the rest — and later refined it even further for efficiency and speed.
This led to an accuracy improvement on real (unseen) data from 62% to 71%.
⸻
Stage 3: Identifying the Best Currency Pairs (+3% Accuracy)
(This stage both helped and hurt us — you’ll see why in a next stage)
At first, we aimed to create a universal neural network trained on all currency pairs, so it could provide predictions across the entire market.
However, during training, we noticed a pattern:
Neural networks trained on just a single currency pair consistently outperformed those trained on multiple pairs.
After analyzing the results, we discovered that:
• Each currency pair has its own unique behavioral patterns
• While some general patterns exist, cross-pair noise reduced accuracy
So we shifted our focus — training separate models for each currency pair.
This boosted accuracy from 71% to 74%.
But here’s the downside:
To support multiple currency pairs now, we must train and maintain a dedicated neural network for each pair, which increases system complexity and resource usage.
So yes — this stage gave us a nice gain in accuracy, but also introduced new challenges.
⸻
Stage 4: Merging Models for a Synergistic Effect (+3% Accuracy)
Here’s the twist:
By chance, we tested a combination of two models:
• One trained on a specific pair
• Another trained on all pairs
We observed that:
• Their combined predictions were less frequent but more accurate.
• The diversity in pattern recognition helped the model filter out noise.
Resulting accuracy: 77%
⸻
️ Stage 5: Data Augmentation, Dropout & Noise (+3% Accuracy)
As the model kept training, we noticed a problem — it began memorizing the data, and accuracy on test sets dropped.
So we implemented:
1. Dropout — randomly disables some neurons during training to prevent overfitting.
2. Input scaling — adds small random variations to inputs, encouraging pattern recognition over memorization.
3. Noise injection — we add slight random noise to candlestick data, training the AI to handle imperfect inputs.
Final accuracy reached: 80% on unseen, real-world test sets.
⸻
What’s Next?
We know that candlestick patterns alone aren’t enough.
The market is influenced by:
• Economic news
• Political events
• Unexpected external shocks
Our next step?
We’re building an AI that will read and analyze financial news in real time, and combine this with our current market pattern model — aiming for even better predictive power.
⸻
Thanks for reading!
We’d love to hear your thoughts, feedback, or suggestions.
Have you tried similar approaches? What worked for you — and what didn’t?
If you’re interested in testing the app, here’s the link:
https://play.google.com/store/apps/details?id=com.PersianDare.AITrading
Let’s build the future of trading together.
