RA in Econometrics. Sorry for the late answer, but I've just read these posts and I found them interesting:

>>22232075

TA is a bit different from models which you use in econometrics, but, if you use 20 years worth of data, consider the idea of implementing time-varying parameters. Indeed, one of the most common problems in finance is balancing the need of having a long time series to fit the parameters and having a sample which belongs to the same population.

As an example, the current market is very different from the one of two years ago. Hence if you were to fit the parameters using two years worth of data, you wouldn't have good predictions.

>>22232410

If, as I understood, the 50% or more relates to the percentage of correct sign predictions then it isn't such a big accomplishment. Personally, I got a percentage of 65% for the US market using dumb models and in “A Recursive Modelling Approach to Predicting UK Stock Returns” by Pesaran & Timmermann (you can download it with SciHub or Libgen) the authors get a percentage of 60%. Moreover, in my projects I also implement what I call the “Chaos benchmark”, aka a vector randomly generated through rand() with suitable adjustments for leverage, and I often get a similar values.

So, in my opinion, he would have to put in more effort to build a truly profitable model.

As for the “no emotion” point, I agree with you. It is certainly a point in favour for algorithmic trading.