Forex Prediction Agent

Project Overview

This project focused on designing an intraday AI prediction agent for the EUR/USD forex pair, combining quantitative market data, technical analysis, and macro-level news signals. The goal was not to “predict the market blindly,” but to explore how AI agents can structure uncertainty, evaluate probability, and support decision-making in fast-moving financial environments.

The system was designed as a research and experimentation platform, emphasizing transparency, evaluation, and risk awareness over raw signal generation.

Problem Statement

Retail trading systems often suffer from:

• Over-reliance on single indicators or black-box models

• Lack of multi-timeframe context

• Ignoring macro news and sentiment

• Poor evaluation and unrealistic backtesting assumptions

I wanted to design a system that treats trading as a decision system, not a prediction toy.

Solution

I built an AI-driven prediction agent that:

• Ingests multi-timeframe MT5 market data

• Combines technical indicators with explicit price-action rules

• Enriches signals with news-based sentiment analysis

• Evaluates predictions through structured backtesting

• Applies risk-aware logic instead of binary buy/sell outputs

The agent produces contextual trade bias and confidence, not blind signals.

System Architecture & Approach

The system follows a modular, research-friendly architecture:

1. Market data ingestion from MT5 across multiple timeframes

2. Feature engineering using technical indicators (RSI, MACD, MAs)

3. Rule-based price-action analysis for structure and confirmation

4. Automated news scraping and sentiment extraction

5. Prediction logic combining technical and macro context

6. Backtesting and evaluation against historical data

7. Risk-aware output (confidence, invalidation, conditions)

Each component was designed to be independently testable and replaceable.

Key Responsibilities

• Designed the intraday prediction architecture for EUR/USD

• Built multi-timeframe market data pipelines from MT5

• Integrated technical indicators and price-action logic

• Automated news scraping and sentiment analysis

• Implemented structured backtesting and evaluation workflows

• Added risk-aware logic to avoid overconfident predictions

• Managed real-time data ingestion and monitoring

Technologies Used

• MetaTrader 5 (MT5) — market data source

• Python — data processing and logic

  • Pandas

  • Technical Analysis (TA) libraries

• n8n — workflow orchestration and automation

• Web Scraping — news ingestion

• OpenAI / Gemini — sentiment extraction and reasoning

Challenges & Learnings

• Indicator stacking adds noise without clear decision structure

• Multi-timeframe alignment is more important than indicator count

• News sentiment improves context but requires careful timing

• Backtesting discipline matters more than model complexity

This project reinforced that most trading failures are evaluation failures.

Outcome & Impact

• Delivered a research-grade prediction agent with transparent logic

• Enabled systematic experimentation across indicators and timeframes

• Provided realistic backtesting and performance evaluation

• Created a reusable framework for future financial AI agents