What Makes a Good Trading Strategy

Dissect the evaluation recipe of a trading strategy

I’ve noticed that many people online sell trading strategies, often showcasing impressive “backtest” results that suggest these methods will perform just as well in the future as they did in the past. While these strategies may appear attractive and promising, a key question arises: how much trust should we place in these backtested results, and how confident can we be in risking our capital based on them? Most importantly, why would someone reveal a highly profitable strategy so easily?

Having developed and analyzed numerous strategies myself, I want to share some insights on how to properly evaluate the performance of a trading strategy and explain why, in many cases, backtested results may not translate to success in real-world trading.

Photo by Traxer on Unsplash

OverFitting

An overfitted strategy is one that appears highly optimized for past data, pinpointing the ideal buy and sell points in historical prices, but struggles to perform consistently in real-world conditions. In essence, overfitting can create the illusion of a perfect backtest, but in reality, it often leads to poor performance and financial losses in actual trading.

For example, suppose I know that a stock price will move from 20 to 40 and then back down to 30. I could adjust my strategy to buy when the price exceeds 20 and sell as soon as it crosses 40. This would result in a flawless profit on paper. However, such precise movements are highly unlikely to repeat in the future, making this strategy unreliable.

So, how do we identify an overfitted strategy? Before deploying or investing in any strategy, it’s essential to consider a few key factors to distinguish genuine robustness from mere curve-fitting.

Understand how the strategy is developed

There are roughly two ways of developing a strategy

1. Idea-driven: Where the initial strategy is coming from an intuitive idea or research paper.
2. Data-driven: When one does not have a very good idea, he/she can leverage some quant strategy development tools, e.g. strategyQuantX to try out different combinations of indicators and find ones that performed well in the past and filter the rest.

Generally, strategies developed through a data-driven approach are more susceptible to overfitting, as they often involve experimenting with various indicator parameters to find the optimal combination. In contrast, strategies based on a well-defined trading idea tend to have a lower risk of overfitting.

When a strategy is derived primarily by “fitting” data, it’s essential to view the results with caution. There’s one idea that might be contrary to many of our beliefs that

A good trading strategy should be simple and intuitive.

Strategies that claim to leverage the most advanced technology and are packed with numerous parameters are often more prone to overfitting past market patterns. In trading, such patterns are less likely to hold up in the future because any “learned” behavior is quickly detected and exploited by other market participants, diminishing its effectiveness over time.

On the other hand, strategies with fewer parameters are less likely to suffer from overfitting. If such a strategy has consistently performed well in the past, it may indicate that it has captured some inherent market dynamics that are more likely to persist.

Enough Years to Test

For a strategy to be considered reliable, it should be validated with at least 10 years of backtested data. The longer the testing period, the less likely it is that the strategy is overfitted. This is because it is significantly more challenging to fine-tune a strategy to perform well over an extended timeframe, particularly when it utilizes only a few parameters.

Effectiveness Across Markets

A good strategy that captures the inevitabilities of markets should work across multiple markets. It is hard to buy in a strategy to be profitable if it works only on Gold. For those kinds of strategies, almost 10 out of 10 are overfitting results.

Breakpoint

To make a short summary here, identifying a good trading strategy that’s not overfitting needs at least these 3 points

1. The strategy should be simple yet intuitive.
2. The strategy should have at least 10 years of testing data.
3. The strategy should work for multiple different markets.

While the rules do not guarantee 100% results, it should be able to filter 8 out of 10 overfitting strategies.

If a strategy passes the 3 conditions, you now have a high chance of having a real deal, but does that mean you can safely believe in the profit curve on the report?

Photo by Sajad Nori on Unsplash

Backtesting Pitfalls

The cruel answer is no, as there are several pitfalls in the backtesting and many times you probably can not believe what you see. Following I will illustrate with the back tester tool from MetaTrader4.

The Accuracy of the Backtest

Most quant strategy testers provide multiple different levels of modeling accuracy. An MT4 example below

It supports 3 types

• Control Points: A faster, less detailed method that approximates tick data using control points within a bar. It uses the high, low, and close prices to estimate intermediate movements.
• Open Prices Only: This method only considers the open price of each bar to evaluate trades, ignoring intraday price movements.
• Every Tick: This is the most precise and detailed backtesting option. It uses every tick data, simulating each individual price movement within a bar.

Based on your trading strategy trigger, if your strategy is sensitive to the price change, you should use tick-level data, while for less sensitive strategies, control point can be used.

Different test accuracy for the same strategy can sometimes produce a very different result. If you’re not clear on which one to go, the rule of thumb is

Always use Tick if your strategy is built for real trading

Modeling Quality

To begin with the importance of modeling quality, let’s look at an example backtest result below

Isn’t this backtest result perfect? The balance curve keeps growing over the whole period with almost no drawdowns. Now check another one,

These two reports look totally different, don't they? What if I tell you these two are using the same strategy with the exact same parameters? Can you figure out what’s gone wrong?

The key difference is at the top right corner Modelling quality , the first graph has a quality n/a , whereas the second has a 99.9% quality.

Most backtesting platforms offer minute-level data as their highest granularity. For any missing data, simulation algorithms are typically used to interpolate price movements. The modeling quality largely reflects how much of the backtest relies on real minute-level data. The higher the modeling quality, the more reliable the testing results, and the closer they reflect actual performance in a live trading environment.

A rule of thumb

A good backtesting result should have a modeling quality of over 90%. Achieving higher levels of accuracy typically requires premium data, as superior data quality comes at an additional cost.

Cost, Cost, and Cost

Cost is very very important in the backtest. If trading were free, even the simplest strategies could perform well in real-world conditions. However, for active trading strategies, the difference between profitability before and after accounting for costs can be significant, often determining whether a strategy is viable in practice.

The general types of costs that must be considered are

1. Commission: The charge on open and close a trade. It generally relates to your position size.
2. Swap: The cost of holding a position overnight. While this cost can sometimes be positive, for most trend-following strategies, this cost can contribute significantly to profitability.
3. Execution cost: Put in other words, slippage cost. For market orders, when you place a trade at $71, you might end up buying/selling the position at $71.1 or $70.9, which means some profit can be taken by the slippage.
4. Spread cost: Some brokers have implicit charges on the spread. You need to relate to the fee compositions to understand how is it charged.
5. Ticket fee: Some brokers also charge a fixed fee when opening a ticket.

To illustrate the importance of cost, if you buy one lot size of USDJPY, which is $100,000, a slippage of 0.001 would result in an extra cost of 0.001 x 100,000 = $100, and the swap can range from $1 ~ 10 based on your holding period, the total cost is not-negligible compare to your profit.

Unfortunately, most backtesting tools do not factor in the real trading cost. The way I follow is to use backtesting to generate trades and use my own simulation system to incorporate real trading costs into the system.

Online Performance

Lastly, once you have everything in place, it’s essential to have a reliable system for deploying and monitoring your trades. It’s important to remember that backtesting does not fully capture future performance. Only in a real trading environment can you truly validate the profitability of your strategy.

Visit us at DataDrivenInvestor.com

Subscribe to DDIntel here.

Join our creator ecosystem here.

DDI Official Telegram Channel: https://t.me/+tafUp6ecEys4YjQ1

Follow us on LinkedIn, Twitter, YouTube, and Facebook.