I have daily time series financial data. I want to apply machine learning techniques to predict expected returns. To do this, I have first transformed the data so that I could take into account time logic into my model as suggested here (link)
I have applied the suggested logic in two different ways after I transformed daily data by dividing daily values with previous day's values and taking their logs (i.e. log(feature1(day2) / feature1(day1))).
1) 5 daily consecutive values of the predictors and expected return values are used as predictors (features) in the model,
2) 5 values of the predictors and the expected return values from previous weeks (i.e. 1 value from last week, another from two weeks ago and so on)
Here is auto-correlation plot for one of the features.
As seen from the figure, first consecutive day has significant correlation. Does this affect my model adversely? When I apply the first logic, I get better results compared to the 2nd. However, when checking the most important features, I always observe the previous days as the most important.
What would be your warnings or suggestions?
