Market Simulation & Multiple Model Voters for better training & prediction performance

Watching one of Jeremy Howard (fast.ai) videos I’ve heard how his student beat the best result of one of Kaggle AI competitions. It was something with texts. But the guy figured out how to increase training data. It is usual practice in machine learning training for images to augment images, rotate, change colors, add noise so on so you would have the same images but different so AI models could learn from way more examples. A cat is still the cat, however, it is rotated in the image. So the idea that is even possible to augment texts made me wonder if the same principles of data augmentation could be applied for my own dataset. One day I got Eureka moment! I can use Fibonacci numbers.  Using this technique now my models can imagine what happens if the price drops 50% or rises 38.2% and so on... Still, it would have to see the same trend whatever numbers we have. The change in data is the same, but the numbers are different. Now my dataset is 6 times bigger, so my models can better learn and “understand” market dynamics..

Another thing I also took from Jeremy, is the model averaging technique. You can train a model with the same data 10 times, but each time it would end up with different weights for neurons due to calculation uncertainty (randomness within the model). So I took it further. I split my dataset into a training/validation set, but each time I do it in different random places. 

It takes all day for long term AI to train one model, I do the training, I see its performance, I send it to my long term AI server. It has a backlog of all models trained before, so it takes the last 12 models, checks each one's performance, and selects the best of 6 models. So next time AI makes predictions, actually 6 models give different predictions and I take the average of that. It is like voting, more voters, more opinions you have of well-trained people, the better final prediction you get. 

For short-term AI I take 24 newest models, evaluate and select 12.
For combined AI, which has the best accuracy we end up with 12 short-term AI votes and 6 long term AI votes. In this combination, we get the best accuracy possible.

And this is how it looks at the backlog of short-term AI trained models with one of the models. Whole brain of the traider fits in the 10 megabytes file.