Freqai or hyperopt?

Which one should I use? Freqai or hyperopt or both? Is freqai better than hyperopt?

Questions above are questions frequently asked by new users. The answer is simple, use whichever can improve your strategy. Both are doing different things and can improve your strategy in different ways. Let me give simple explanation about what each of them actually does.

Hyperopt

Hyperopt is just series of backtests done to find the param(s) that can produce lowest loss value. It uses historical data to find those param(s). There is no guarantee that param(s) that works well in the past will still perform good in the future. Hyperopt can’t predict future since there is no machine learning involved in it.

Freqtrade won’t run hyperopt process automatically. So if you want to update the param(s) in the future, either you have to manually run the hyperopt process or you write a script to run it automatically on specified interval.

Example of hyperopt is you want to find which EMA is good for your dataframe["close"] < dataframe["EMA_x"] logic. Based of your specified loss function, let’s say hyperopt found that EMA_100 is the best param. Will it still perform good in the future? No one knows.

Freqai

For simplicity, I will describe what regressor/classifier model does in this article. In simple term, what they do is it takes defined features (which is based of past data) and use those data to predict specified (future) targets. Freqai don’t know (and don’t care) whether the strategy is profitable or not. All it cares is to produce good predictions. It’s your responsibility to know how to use the targets to create profitable strategy.

For example. let’s take this official example strategy. The model will use the defined features (rsi, mfi, hour, etc) to predict the target dataframe["&-s_close"], which is predicting the average movement of price label_period_candles candles in the future. How you use the prediction is up to your creativity. The logic used in the example may or may not be profitable. But as said above, that isn’t the focus of freqai.

The advantage of using freqai is it will re-train as often as defined, which means it will adapt automatically to the market by re-calculating the target(s). The downside of freqai is as I have repeated multiple times above, there is no guarantee that even if it can predict the target(s) accurately, it would make your strategy profitable. Not even any guarantee of past profit, like what hyperopt at least can offer (not that past profit guarantee future profit).

So which one is better?

Neither are better than other. Both do different things and focus on different aspects. Personally, I’m using both on some of my strategies. You just need to understand what each of them offers and how to utilize them correctly. From my experience, freqai might take more time to be learned than hyperopt, because you will need many trial-and-errors to find good combination of features and targets. You don’t want too much features that it would take ages to train the model. Some features might not be good features or redundant to similar existing features.