Hi All, Having read this excellent article: https://www.analyticsvidhya.com/blog/2018/10/predicting-stock-price-machine-learningnd-deep-learning-techniques-python/
And many more like it on various sites that focus on pretty much the same method, I have tried a number of experiments, using Tensorflow, AutoSKLearn and even Googles AutoMLTables and they all seem to suffer from the same problem - namely an inability to predict with enough accuracy to make consistently profitable trades on.
Some research seems to indicate that this type of model suffers from a Random Walk problem, in that each prediction uses all data up to the previous time period to make the prediction for the next time period. Then when it predicts the next and subsequent time periods values, the prediction is influenced by the prior period’s values and hence the graph line “looks” like it fits well. In reality all that is happening is that a value very similar to the last value is being predicted, hence we end up with a prediction that tells us what the price has been and not what it is going to be.
Zooming in closely on this plotted data shows the inaccuracy of the predictions quite well and setting up a backtest for actual trades shows that the profit / loss is no better than a buy and hold strategy.
I have tried price prediction on different time frames (from 1m to 4h), multivariate input including OHLCV plus a number of calculated indicator features, classification of 60 minutes worth of data into up or down, in an attempt to predict the future hour’s direction and a few other things, but nothing gets better than market returns on backtesting.
So, my question is, has anyone managed to create any models or identify any techniques that provide better than market returns consistently?