How similar are follow through bars to previous bars?

This study attempts to prove or disprove a hypothesis that a bar's appearance affects the bar that comes next. It compares all pairs of consecutive bars in a very large dataset and assigns a similarity score to all the follow through bars.

Visually the question asks if regardless of whether a bar is a trend bar or a doji, to what extent will the follow through bar look similar (not necessarily exact but close)?

Methodology

The similarity function uses the three main features of a bar: upper tail, lower tail and body. It then calculates percentage difference between two given bars for each feature separately. For example, if the upper tails on each of the bars are exactly the same size it scores the tails at 100%. If one bar has a tail and the other one doesn't it scores that feature as 0%. Similar process is repeated for bodies and lower tails. If the follow through bar is of an opposite color to the previous bar then the whole process defaults to 0% and comparison moves on to the next pair.

It's tempting to just average the scores now and be done with it, but there's a catch: since the similarity function defaults to 0% as soon as one of the bars doesn't have a tail and the other one does, it creates a situation where two pairs of different looking bars end up with the same score, which doesn't seem correct.

A workaround is to first evaluate the sizes of each feature relative to the range of a bar. That way, if the tails are small compared to the rest of the bar they won't weigh as much on the final result.

And now for the final magic trick, we use the size comparisons calculated at the very beginning to offset these weights.

That was easy. We now have our similarity score function working and can run it against the historic dataset.

Results

Now comes the fun part. This process was run on over 337K bars so any context that may have influenced follow through bar behavior would've been canceled out. We'll use histograms to analyse the results. If the hypothesis is correct then the majority of follow through bars will cluster around the upper range of the similarity function.

Result 1

Oh no! What's going on? There is a huge spike on the left completely offsetting our beautiful similarity function results. Well, remember that if a follow through bar is of an opposite color the function defaults to scoring it as simply 0. This result tells us that majority of FT bars are actually more likely to be of opposite color. There were about 200K FT bars of opposite color. Since this is a historgram, we sum up the rest of the similarity scores that fall above 0 to give us a count of how many FT bars were of same color: 137K. This means that there's only a 40% chance to even get an FT bar of same color (excluding any other context).

Result 2

Let's isolate just the results greater than 0 and focus on the FT bars that were of the same color. This result is just as sad. The majority of FT bars of same color cluster around a "half similar" region. What we were hoping to see, clustering around the upper range of similarity (>0.80 and above) constitutes a very small percentage of the total.

Conclusion

This empirically disproves the hypothesis that any given bar, taken out of context, has any influence on the follow through bar. As a word of caution to anyone wanting to use this result to build a countertrend strategy: this result doesn't say anything about reversals or strength of moves in the opposite direction. A strong bull bar followed by an opposite color bar could be just a bearish doji and the trend may still resume higher. This study doesn't cover any of that, only what or more precisely what not to expect. As always, traders are encouraged to replicate and test these results. All the code is freely available on GitHub.