We have started a semi-AI trading plan to recover the money lost in Forex.
--The foreign exchange market is not a linear problem. Only simple parameters can be defined --If there are few explanatory variables, you can never make a highly versatile model. —— Due to the nature of the market, the overall market environment is different, so the same input can produce different results. ――This is something that computers are not particularly good at --Computers cannot understand abstract concepts like "behavior" and "market" --Even computers that play world-class chess don't even know what the actual chess is. --So the computer just knows how to win --Time-dependent and non-deterministic exchange rates are difficult for machine learning systems --You need to understand the market situation essentially ――Live trading is quite difficult considering the time required for learning and verification. --It doesn't make sense to simply build a model based on your ability to make accurate predictions (up or down) --99 small decisions can be wiped out by one big decision --To avoid this problem, use a method of retraining machine learning algorithms.
① By using a flexible window, the selection bias associated with the in-sample / out-of-sample set is reduced. (2) The training will be a series of verification exercises so that the machine learning algorithm will always function properly even if the data sets are significantly different. In the process of establishing the true value and method of an algorithm, measuring backtesting performance should also be considered important.
――This approach is not fail-safe (always safe control) --Biased curve fitting (overfitting) and data mining (knowledge extraction technology) --To mitigate this, we used a huge dataset, performed a large number of data mining bias evaluation tests, and It is important to eliminate the possibility of accidental promising results (creating a semi-perfect market situation algorithm)
③ Based on the above, can you grasp the current market in real time and increase the chances of success with customizable algorithms? May
Currently under consideration, conclusions will be drawn soon
Recommended Posts