@Arindam Bhattacharjee Good to see you are using Data analytics to make predictive analysis. This is an upcoming field and we are already part of AI based ecosystem and will see lot more in near future.
Few comments based on the information you have shared and for understanding of others:
1. Any particular reason you have used KNN (K-Nearest Neighbor) Algorithm? As this algorithm is more suitable to solve classification problems. For example, in predictive recommendation systems such as those used in online portals (Amazon, Flipkart etc) where you see "
you may also like this" kind of recommendations.
2. I presume you are using this algorithm to 1st predict only the probable direction the stock will move based on the 6-7 parameters and then use some regression algorithm to measure the probable extent of that move?
3. Just a suggestion, you may explore using simple ensemble models of Random Forrest along with linear regression algorithm. I have observed they give better results.
4. These ML models just provide static analysis, you can try making it more interesting & efficient by using Reinforcement learning (Deep learning) techniques on top to make it a true AI application which will learn dynamically from its mistakes and successes and take intelligent calls (Similar to how computers play Chess/Games these days and there are lot of research papers available on this topic).
However word of caution based on my experience:
1. Strategies based on these type of predictive models will work best if they are fully automated else if human intervention is involved, he will most likely second guess the trade signals when the strategy enters draw-down period. Hence the models should be clear to understand and not completely black box so that one has confidence in using it.
2. Each ML model is a strategy based on machine learned rules, it does not mean it will not fail in future. It is just like we running one of our successful strategies mechanically, but there will be periods/conditions after which the strategies stop working and we many need to make changes to it. Similarly ML based models/strategies also require regular monitoring/refining/optimization.
3. The edge in these types of models are very small, i.e it cannot scale well and if it is posted on the open forum the edge will be lost much faster. I remember testing a strategy mentioned by
@UberMachine in his automation thread and it works but the edge is very small and it cannot scale up well if many trade it and overhead costs are included.
4. Hence best way to apply machine learning based models is to diversify it, i.e select multiple strategies (5-10) which are not co-related and run them all in parallel with proper MM & Risk control.
ML based strategies works but be clear that it will not make you rich overnight and the results will be inline with index returns but with some additional alpha based on the model's edge used in these strategies.
Frankly at the moment I think it is more suitable for large fund management running hundreds of strategies in parallel, for poor retailers like us using our normal cognitive intelligence to trade mechanically should suffice...
Also no fun being on the sideline and watch computers having all the excitement right?