MACD Crossover Strategy Backtesting Across vectorbt, Nautilus Trader, and MetaTrader 5

Project Title

MACD Crossover Strategy Backtesting Across vectorbt, Nautilus Trader, and MetaTrader 5

Project Overview

This project implements the same manually calculated MACD crossover strategy across three backtesting and trading environments:

• vectorbt
• Nautilus Trader
• MetaTrader 5

The goal is to compare implementation behavior across a vectorized Python framework, an event-driven Python trading engine, and the MetaTrader 5 Strategy Tester. The project prioritizes clear assumptions, reusable shared Python logic, and honest documentation of framework differences.

Submission Summary

• This repository contains the complete take-home assignment submission.
• The same MACD crossover strategy was implemented across vectorbt, Nautilus Trader, and MetaTrader 5.
• MACD and EMA calculations were implemented manually.
• EUR/USD H1 data for 2024 was used.
• Final results and the MT5 exported report are included.
• Repository URL: https://github.com/Sigmanumeric78/MACD-crossover.git

Strategy Definition

• Instrument: EUR/USD
• Timeframe: H1
• Date range: 2024-01-01 to 2024-12-31
• Initial capital: 100000 USD
• Python trade size: 10000 EUR/USD units
• MT5 trade size: 0.10 lot
• Mode: long-only / flat
• Fast EMA period: 12
• Slow EMA period: 26
• Signal EMA period: 9
• MACD line: fast EMA minus slow EMA
• Signal line: EMA of the MACD line
• Entry: enter long when MACD crosses above the signal line
• Exit: exit to flat when MACD crosses below the signal line

The EMA uses SMA seeding: the first EMA value appears after period valid values and is seeded with the simple average of those values. Later EMA values use alpha = 2 / (period + 1).

No built-in MACD or EMA indicators are used. The Python implementations avoid pandas .ewm(), TA-Lib, pandas-ta, vectorbt indicators, and Nautilus indicators. The MT5 Expert Advisor avoids built-in indicator handles and external indicator files.

Signals are generated after candle close. Execution is intended on the next bar open, or the equivalent event after a completed bar is received.

Data Source and Preparation

The real dataset is ForexSB Historical Data EURUSD60.csv, normalized into the project schema.

• Raw source file: data/raw/EURUSD60.csv
• Processed Python file: data/processed/EURUSD_H1_2024_clean.csv
• MT5 import file: data/mt5_import/EURUSD_CUSTOM_H1_2024.csv
• Processed rows: 6250
• Processed date range: 2024-01-01 22:00:00 to 2024-12-31 21:00:00
• Processed columns: timestamp, open, high, low, close, volume

Weekend and holiday gaps are normal for forex H1 data. The most common processed timestamp gap is 60 minutes, while larger gaps appear around market closures.

Prepare the processed CSV from the raw ForexSB file:

python data/prepare_eurusd_h1_data.py --source csv --input data/raw/EURUSD60.csv

The MT5 import CSV uses:

Date, Time, Open, High, Low, Close, TickVolume, Volume, Spread

No timestamp shifting is applied. Project timestamps are treated as UTC/no-shift bar open times.

Repository Structure

macd-backtest-test/
├── README.md
├── MEMORY.md
├── .gitignore
├── requirements-vectorbt.txt
├── requirements-nautilus.txt
├── common/
│   ├── config.py
│   ├── macd.py
│   ├── data_loader.py
│   ├── signals.py
│   └── metrics.py
├── data/
│   ├── raw/EURUSD60.csv
│   ├── processed/EURUSD_H1_2024_clean.csv
│   ├── mt5_import/EURUSD_CUSTOM_H1_2024.csv
│   └── prepare_eurusd_h1_data.py
├── vectorbt_impl/
│   ├── run_vectorbt_backtest.py
│   ├── vectorbt_results.csv
│   └── vectorbt_trades.csv
├── nautilus_impl/
│   ├── macd_strategy.py
│   ├── run_nautilus_backtest.py
│   ├── nautilus_results.csv
│   └── nautilus_trades.csv
├── mt5_impl/
│   ├── MACD_Crossover_Manual.mq5
│   └── reports/
│       ├── mt5_report.html
│       └── mt5_results_summary.md
└── reports/
    ├── vectorbt_equity_curve.csv
    └── nautilus_equity_curve.csv

Environment and Dependencies

Install vectorbt dependencies:

pip install -r requirements-vectorbt.txt

Install Nautilus Trader dependencies:

pip install -r requirements-nautilus.txt

This project was developed with Python 3.13 in the local environment. vectorbt required NUMBA_DISABLE_JIT=1 to avoid a numba cache import issue in this environment; the vectorbt runner sets this before importing vectorbt.

MetaTrader 5 requires a separate MT5/MetaEditor installation. The Strategy Tester run was performed manually using the exported Expert Advisor and custom-symbol CSV.

How to Run

Run vectorbt:

python vectorbt_impl/run_vectorbt_backtest.py

Run Nautilus Trader:

python nautilus_impl/run_nautilus_backtest.py

Run MT5 manually:

1. Compile mt5_impl/MACD_Crossover_Manual.mq5 in MetaEditor.
2. Import data/mt5_import/EURUSD_CUSTOM_H1_2024.csv into a custom symbol such as EURUSD_CUSTOM.
3. Run Strategy Tester:
   • Expert: MACD_Crossover_Manual
   • Symbol: EURUSD_CUSTOM
   • Timeframe: H1
   • Date range: 2024-01-01 to 2024-12-31
   • Initial deposit: 100000 USD
   • Lot size input: 0.10
   • Modeling mode: Open prices only
4. Save the MT5 report to mt5_impl/reports/mt5_report.html.
5. Record the manual MT5 results in mt5_impl/reports/mt5_results_summary.md.

Results Summary

Framework	Data	Trades	Total Return	Sharpe Ratio	Max Drawdown	Notes
vectorbt	REAL_CSV	241	-0.3863%	-0.9415	-0.5761%	Vectorized signal portfolio
Nautilus Trader	REAL_CSV	241	-0.4280%	-6.1039	-0.5740%	Event-driven; equity reconstructed from closed PnL
MetaTrader 5	EURUSD_CUSTOM	230	-0.4325%	-1.23	-0.65%	Manual Strategy Tester run; 16% history quality and 6106 bars

These results are not profitable. They are included to compare implementations and execution behavior across frameworks.

Framework-by-Framework Notes

vectorbt:

• Uses shared manual MACD logic from common/macd.py.
• Uses vectorized entries/exits from common/signals.py.
• Shifts signals by one bar and uses open prices for execution.
• Writes vectorbt_impl/vectorbt_results.csv, vectorbt_impl/vectorbt_trades.csv, and reports/vectorbt_equity_curve.csv.

Nautilus Trader:

• Implements manual EMA/MACD inside an event-driven Strategy subclass.
• Uses completed H1 bar events and submits market orders through a simulated venue.
• Uses a margin account for EUR/USD because the installed Nautilus API rejected CurrencyPair instruments on a single-currency cash account.
• Writes nautilus_impl/nautilus_results.csv, nautilus_impl/nautilus_trades.csv, and reports/nautilus_equity_curve.csv.
• The Nautilus equity curve is reconstructed from closed position PnL because a full bar-level equity curve was not exposed by the low-level report API used here.

MetaTrader 5:

• Uses mt5_impl/MACD_Crossover_Manual.mq5.
• Calculates EMA/MACD manually in MQL5.
• Runs through MT5 Strategy Tester on EURUSD_CUSTOM.
• Report is saved at mt5_impl/reports/mt5_report.html.
• No screenshot file is included.

Why Results Differ

The implementations intentionally share the same strategy rules, but the frameworks execute and report differently:

• vectorbt uses vectorized signal arrays and shifted boolean masks.
• Nautilus Trader is event-driven and routes orders through a simulated trading venue.
• MetaTrader 5 uses Strategy Tester, custom-symbol processing, and its own order/fill engine.
• MT5 reported 6106 bars and 230 trades, while the Python processed CSV has 6250 rows and both Python frameworks report 241 trades.
• MT5 also reported History Quality of 16%, which limits direct numerical comparability.

The goal is implementation comparison, not perfect bit-level replication.

Final Deliverables

• Source code for the vectorbt implementation.
• Source code for the Nautilus Trader implementation.
• MQL5 Expert Advisor for MetaTrader 5.
• Cleaned EUR/USD H1 dataset: data/processed/EURUSD_H1_2024_clean.csv.
• MT5 import CSV: data/mt5_import/EURUSD_CUSTOM_H1_2024.csv.
• vectorbt result CSVs: vectorbt_impl/vectorbt_results.csv and vectorbt_impl/vectorbt_trades.csv.
• Nautilus result CSVs: nautilus_impl/nautilus_results.csv and nautilus_impl/nautilus_trades.csv.
• MT5 exported HTML Strategy Tester report: mt5_impl/reports/mt5_report.html.
• MT5 result summary: mt5_impl/reports/mt5_results_summary.md.
• Root README and framework-specific README files.
• Development log: MEMORY.md.
• PDF submission summary: reports/MACD_Backtesting_Project_Summary.pdf.

AI Tool Usage Notes

• AI coding tools were used as development support for scaffolding, implementation planning, debugging, documentation drafting, and cross-framework reasoning.
• All major outputs were verified through actual terminal runs for vectorbt and Nautilus Trader and through a manually exported MT5 Strategy Tester report.
• AI-generated suggestions were validated against runtime behavior and corrected where necessary.
• Development decisions, commands, errors, fixes, and unresolved issues are recorded in MEMORY.md.

Final Notes

The repository contains the completed implementation, generated reports, and documentation for submission. The current result files are small enough to keep in the repository for this take-home style project. The main remaining limitation is that MT5 custom-symbol history quality and bar count differ from the Python processed dataset, so cross-framework comparisons should be interpreted as implementation-level comparisons rather than exact numerical replication.