The SPX model accuracy (since 2010 test set) for correctly predicting positive/negative days:
The predictors I used were: VIX, Crude, Gold and Previous Day’s Close (factor).
The results were factors either returning 1 or 0 if the following day’s close was higher than the previous.
The train/test data set start date is 2007-12-31. I first tried data partitioning (75/25) the train and test sets in a random fashion irrespective of time-series. Then I tried testing it in slices (75/25) so time was continuous. Results were very little changed.
Here is how the SVM model performed using those 4 variables:
The results were fairing unimpressive but better than 50/50 is a great start.