This model excludes Tuesday data from 7.5 as it preceded FOMC minute publication and occurred after a 3 day holiday weekend.
Look to take advantage of early morning market, with possible pump followed by a drop. Try to exit before 11:15 where historical chops occur. Best to re-enter market after 2:45.
Patterns are lagging. The pattern can hold true whether the market closes up or down.
Look to take advantage of early morning market, with possible pump followed by a drop. Try to exit before 11:15 where historical chops occur. Best to re-enter market after 2:45.
Patterns are lagging. The pattern can hold true whether the market closes up or down.
Trade closed manually:
The model held up well until the 230 swing point where the pattern sharply reversed. I have updated the model and calculated a deviation score. Looking at a straight deviation score of 1.41% shows that the model captured most of the day's moves. However, when weighting the deviation score by volume, we get a miserable -18.59%. We want to be within +/- 10% deviation. So this model is currently undergoing a transition, and will eventually catch up with the market. But at this time has higher unreliability--at least in the afternoon sessions.
Trade closed manually:
So I have done even more calculating and realized that it simply does not make sense to weight deviation by volume. I had tried weighing the intra-day patterns by volume, but it skewed the entire patterns to something unrecognizable. In short, there is so much volume the first 30 minutes and last 30 minutes, it essentially would degrade any intra-day modeling across the day. It's best to keep with a moving average. With that in mind, the same is holding true for the Deviation Score. So I am officially updating this model to an Deviation Score of 1.59% - an outstanding result. I also found an error in my weighted deviation score, so if I were to weight it by volume, it improves by 5% up to -14.94%. All this means is that if I were to look at the first 30 minutes on a Tuesday, The actual forecasted value was much higher than the model predicted---but the overall pattern/curvature remained the same. I may need to do some more advanced research to calculate curvature (logs) and compare.
Comment:
I ran a couple of statistical analysis on the models (K-S test of dissimilarity) and a coefficient of determination based on a signal analysis methodology. Both models showed that Tuesday's model was a better match than Monday's. I'll be using these different methods to cross check my deviation scores and compare models going forward.
Comment:
Per K-S Test, 6% dissimilarity. I.E., 94% similar. This test was run against the closing prices on each 5 minute interval.