I've been experimenting with this a lot. My code is all public still because I haven't made any giant gains or anything. But I have seen some success here and there. I've even got this one bot that learns from its past trades via ML and uses what it has learned to decide wether to make future trades or not. https://github.com/madchops1/Dutchess.ai
I apologize. I can do that. I'm going to generate that backtest you described.
Right now I have residual data from the AWS machine learning data that tells me weather there is any structure to the times it does guess wrong. And a value below baseline is a better than 50/50 guess according to what I have learned about how AWS does its ML. Knowing that I use this personally as a supporting indicator to my trade decisions. Since its so new and I really don't want people to think I'm scamming or something. I'm just releasing my results free for now, not trying to be a douche ;)
AWS defines the baseline as follows
Baseline RMSE
Amazon ML provides a baseline metric for regression models. It is the RMSE for a hypothetical regression model that would always predict the mean of the target as the answer. For example, if you were predicting the age of a house buyer and the mean age for the observations in your training data was 35, the baseline model would always predict the answer as 35. You would compare your ML model against this baseline to validate if your ML model is better than a ML model that predicts this constant answer.
I did another back test with a randomly selected 70%/30% training to evaluation ratio for evaluating time-series models. Adding results to whitepaper. The results are still under baseline.
Basically I have a dataset and I train my model with 70% and then evaluate its guesses against the remaining 30%. Hence a baseline is created and I can see if my model performs better.
It took some doing to get this model to perform well. I did this by adding features that help recognize patterns in the time series data.
The features I created are not specific to QM as they are technical (eg. numbers, not news), and time-series related. So the models should work with any historical dataset with the same fields.
I am not trading purely on these numbers. They are indicators they help me make my trading decisions and I think they may help others too.
My models evaluations are performing better than baseline when trained with 70% of the data and evaluated against the remaining 30% so I take that as value. As someone else put it a potentially "favorable guess". At this point I'm using the predictions regularly. And I guess I'll know more the longer I keep track of daily results.
I experimented with the features until I got good results in my evaluations. If I am getting favorable guesses, how does that point to bad features. Favorable guesses is what I was going for in order to assist in trading decisions.
I published the evaluation results in the whitepaper. I am doing my training/evaluation with a data split of 70/30. Doesn't that qualify as a proper backtest?
Your are correct about no hedge agains outlying macro events. But I'm not suggesting using this without any other variables in your trading decision. It may help make a trading decision.