By Ajay Pawar
This blog serves as an introduction to key concepts, but for a solid foundation in backtesting, it’s recommended to start with How to Backtest: Strategy, Analysis, and More. This will help you understand the fundamentals of testing, and analysing trading strategies before deploying them.
Introduction
Traditional backtesting assumes that optimising a strategy on historical data and validating it on an out-of-sample period ensures future reliability. Traders typically backtest on in-sample data, optimise parameters, and validate on a brief out-of-sample period. If results look good, they assume robustness and move to live trading.
Figure: Static Backtesting Method
However, this approach has some drawbacks. Overfitting is a significant problem—strategies often reflect past patterns rather than being robust. If a strategy is optimised on most historical data, leaving only a small portion for validation, this limited test can give false confidence.
Additionally, validation over the static out-of-sample period with fixed parameters fails to account for changing market conditions. Traditional backtesting is also static, assuming fixed parameters remain effective despite ever-changing markets. In reality, traders constantly reassess strategies based on recent data. Without ongoing validation, static backtesting creates an overfit, backwards-looking view, offering little assurance of future success. Traditional backtesting does not reflect real-world performance. A strategy that appears profitable in backtests may collapse in live trading because fixed-period validation does not test its ability to adapt to new data.
This blog covers:
- Implementing Backtesting with Walk Forward Optimisation (WFO) Framework
- Why Walk-Forward Optimisation (WFO) Works
- Why Walk-Forward Optimisation (WFO) Works
- Limitations of Walk-Forward Optimisation (WFO)
Implementing Backtesting with Walk Forward Optimisation (WFO) Framework
These issues can be addressed by implementing backtesting with the WFO framework, which continuously re-optimises strategy parameters using a rolling-window approach. Instead of a single optimisation-validation split, WFO cycles through multiple periods, progressively incorporating new data while testing on unseen market conditions.
How to Implement backtest with WFO:
Figure: Backtesting with Walkforward Optimisation.
Consider a portfolio allocation strategy using price data from 2010 to 2025. Instead of the traditional approach—optimising weights using data from 2010-2021 and testing on 2022-2025—WFO creates a series of optimisation-validation cycles:
In this strategy, we will implement Walk-Forward Optimization (WFO) using a rolling in-sample window of the past five years to optimise strategy parameters. These optimised parameters will then be applied to an out-of-sample window spanning one year, enabling us to assess the strategy's performance on the subsequent year's data.
Steps:
- Initial Cycle: Optimise portfolio weights using data from 2010-2015 (in-sample period).
- First Validation: Apply these optimised weights to 2016 and record performance (out-of-sample).
- Rolling Forward: Shift the in-sample window forward by one year (now 2011-2016).
- Second Optimisation: Re-optimise weights using this updated in-sample data.
- Second Validation: Apply these newly optimised weights to 2017.
- Continuous Process: Repeat this process, maintaining a consistent in-sample window length, until reaching 2025.
This creates a sequence of optimisation-validation pairs:
- Train on 2011-2015, test on 2016
- Train on 2012-2016, test on 2017
- Train on 2013-2017, test on 2018
- And so on through 2025
By combining these out-of-sample period results, we create a more realistic assessment of how the strategy would have performed if traded throughout this period.
This approach is also highly relevant when using ML-based trading strategies. By training on an evolving in-sample period and validating on a rolling out-of-sample window, ML-driven trading models can mitigate overfitting and improve their ability to generalise to new market conditions.
Why Walk-Forward Optimisation (WFO) Works
Walk-forward optimisation reduces overfitting by testing each segment of data in a forward-looking manner, preventing the false confidence that can come from a single, potentially lucky validation period. Your strategy must prove itself repeatedly across different market conditions, creating a more rigorous validation process.
Unlike traditional backtesting, which assumes parameters remain effective indefinitely, WFO reflects how traders actually operate—continually reassessing and adjusting strategy parameters as new market data becomes available. This creates a dynamic approach that better mimics real-world trading behaviour. Additionally, WFO maximises data efficiency since each time period serves dual purposes: first as an out-of-sample validation period, then as part of the next in-sample optimisation window. This means most of your historical data contributes to both training and testing.
Limitations of Walk-Forward Optimisation (WFO)
Despite these advantages, Walk-Forward Optimisation isn't without important limitations.
Window Selection Bias: The size of your training and testing windows fundamentally shapes your results. Too short a training window misses essential market cycles and produces unstable parameters, while too long a window incorporates outdated market conditions that may no longer be relevant. Even the specific starting points of your windows can capture seasonal effects or unique market periods that skew results, creating another source of potential bias in your testing framework.
Market Regime Changes: While WFO adapts better than static backtesting, it still responds to regime changes with a lag. When markets transition between major states such as bull markets, bear crashes, or sideways consolidation, strategy performance often deteriorates before WFO can adjust the parameters appropriately. You're still discovering regime shifts after experiencing their negative impacts, which means the approach maintains a reactive rather than truly predictive character.
Computational Complexity: The repeated re-optimisation process increases computational demands. Compared to a single backtest, WFO requires multiple rounds of optimisation and validation, making it resource-intensive, especially for complex or high-frequency strategies.
Conclusion
Traditional backtesting is limited by its static nature and susceptibility to overfitting, making it an unreliable predictor of future performance. Walk-forward optimisation (WFO) offers a more adaptive alternative, continually reassessing strategy parameters through a rolling-window approach. By validating performance across multiple out-of-sample periods, WFO better simulates real-world trading.
However, WFO has limitations. Window size selection impacts results, introducing biases, and while it adapts to market changes, it reacts to regime shifts rather than predicting them. Its computational demands also pose challenges, especially for complex or high-frequency strategies.
Despite these drawbacks, WFO enhances traditional backtesting by providing a more rigorous, adaptive validation process, increasing the likelihood of sustained profitability.
Next steps
Once you’re comfortable with the basics, explore advanced techniques to enhance your backtesting and strategy validation process.
If you want to read our Quantra Classroom about WFO and how to implement it with LSTM, you will learn how to realistically backtest the LSTM neural network created to calculate the optimum weights of assets in the portfolio using the walk-forward optimisation method.
For Python implementation of the walk-forward optimisation method for the LSTM network in the context of portfolio management, check out the Quantra Course Section: Walk Forward Optimisation With LSTM.
The next step in this series will introduce Walk-Forward Optimization (WFO) using XGBoost, providing a structured way to optimize and validate trading models dynamically.
Cross-Validation for Model Testing
Cross-validation plays a crucial role in ensuring that trading models are not overfitting to historical data. Learn how to refine model validation using advanced techniques:
- Cross-Validation: Embargo, Purging & Combinatorial Methods – Explore techniques to remove biases and improve model reliability.
- Cross-Validation for Machine Learning-Based Trading Models – Understand how different cross-validation techniques can be applied to machine learning-driven trading models.
Structured Learning
To gain hands-on experience in backtesting and validation, consider these resources:
- Backtesting Trading Strategies (Quantra Course) – A structured course covering strategy backtesting, optimization, and performance evaluation.
Backtesting with Blueshift
- Blueshift – Institutional-Grade Backtesting & Trading Platform – An all-in-one platform designed for research, backtesting, and algorithmic trading. Blueshift provides a fast, flexible, and reliable solution for testing strategies across various asset classes and trading styles.
