A few important steps for creating robust and winning StrategiesAs the title says, I want to share knowledge & important insights into the best practices for creating robust, trustworthy and profitable trading Strategies here on TradingView.
These bits of information that my team I have gathered throughout the years and have managed to learn through mostly trial and error. Costly errors too .
Many of these points more professional traders know, however, there are some that are quite innovative for all levels of experience in my opinion. Please, feel free to correct me or add more in the comments.
There are a few strategic and tactical changes to our process that made a noticeable difference in the quality of Strategies and Indicators immediately.
Firstly and most importantly, we have all heard about it, but it is having the most data available. A good algorithm, when being built NEEDS to have as many market situations in its training data as possible. Choppy markets, uptrends, downtrends, fakeouts, manipulations - all of these are necessary for the strategy to learn the possible market conditions as much as possible and be prepared for trading on unknown data.
Many may have heard the phrase "History doesn't repeat itself but rhymes well" - you need to have the whole dictionary of price movements to be able to spot when it rhymes and act accordingly.
The TradingView Ultimate plan offers the most data in terms of historical candles and is best suited for creating robust strategies.
___
Secondly, of course, robustness tests. Your algorithm can perform amazingly on training data, but start losing immediately in real time, even if you have trained it on decades of data.
These include Monte-carlo simulations to see best and worst scenarios during the training period. Tests also include the fundamentally important out-of-sample checks . For those who aren’t familiar - this means that you should separate data into training sets and testing sets. You should train your algorithm on some data, then perform a test on unknown to the optimization process data. It's common practice to separate data as 20% training / 20% unknown / 20% training etc. to build a data set that will show how your algorithm performs on unknown to it market movements. Out of sample tests are crucial and you can never trust a strategy that has not been through them.
Walk-forward simulations are similar - you train your algorithm on X amount of data and simulate real-time price feeds and monitor how it performs. You can use the Replay function of TradingView to do walk-forward tests!
When you are doing robustness tests, we have found that a stable strategy performs around 90% similarly in terms of win rate and Sortino ratio compared to training data. The higher the correlation between training performance and out of sample performance, the more risk you can allocate to this algorithm.
___
Now lets move onto some more niche details. Markets don’t behave the same when they are trending downward and when they are trending upwards. We have found that separating parameters for optimization into two - for long and for short - independent of each other, has greatly improved performance and also stability.
Logically it is obvious when you look at market movements. In our case, with cryptocurrencies, there is a clear difference between the duration and intensity of “dumps” and “pumps”. This is normal, since the psychology of traders is different during bearish and bullish periods. Yes, introducing double the amount of parameters into an algorithm, once for long, once for short, can carry the risk of overfitting since the better the optimizer (manual or not), the better the values will be adjusted to fit training data. But if you apply the robustness tests mentioned above, you will find that performance is greatly increased by simply splitting trade logic between long and short. Same goes for indicators.
Some indicators are great for uptrends but not for downtrends. Why have conditions for short positions that include indicators that are great for longs but suck at shorting, when you can use ones that perform better in the given context?
___
Moving on - while overfitting is the main worry when making an algorithm, underoptimization as a result of fear of overfitting is a big threat too . You need to find the right balance by using robustness tests. In the beginning, we had limited access to software to test our strategies out of sample and we found out that we were underoptimizing because we were scared of overfitting, while in reality we were just holding back the performance out of fear. Whats worse is we attributed the losses in live trading to what we thought was overfitting, while in reality we were handicapping the algorithm out of fear.
___
Finally, and this relates to trading in general too, we put in place very strict rules and guidelines on what indicators to use in combination with others and what their parameter range is. We went right to theory and capped the values for each indicator to be within the predefined limits.
A simple example is MACD . Your optimizer might make a condition that includes MACD with a fast length of 200, slow length of 160 and signal length of 100. This may look amazing on backtesting and may work for a bit on live testing, but these values are FUNDAMENTALLY wrong (Investopedia, MACD). You must know what each indicator does and how it calculates its values. Having a fast length bigger than the slow one is completely backwards, but the results may show otherwise.
When you optimize any strategy, manually or with the help of a software, be mindful of the theory. Mathematical formulas don’t care about the indicator’s logic, only about the best combination of numbers to reach the goal you are optimizing for - be it % Return, Profit Factor or other.
Parabolic SAR is another one - you can optimize values like 0.267; 0.001; 0.7899 or the sort and have great performance on backtesting. This, however, is completely wrong when you look into the indicator and it’s default values (Investopedia, Parabolic SAR).
To prevent overfitting and ensure a stable profitability over time, make sure that all parameters are within their theoretical limits and constraints, ideally very close to their default values.
Thank you for reading this long essay and I hope that at least some of our experience will help you in the future. We have suffered greatly due to things like not following trading theory and leaving it all up to pure mathematical optimization, which is ignorant of the principles of the indicators. The separation between Long / Short logic was also an amazing instant improvement.
View the linked idea where we explain the psychology of risk management and suggest a few great ways to calculate and manage your risk when trading - just as important as the strategy itself!
What do you think? Do you use any of these methods; Or better ones?
Let us know in the comments.
