Predicting Return on Investment for Financial Data Using Different Machine Learning Modeling Techniques
The purpose of our project is to investigate the effectiveness of various machine learning techniques and algorithms for predicting an investment’s rate of return based only on historical market data. Specifically, our goal is to determine the effects that different modeling techniques have on the performance of our end model. Therefore, we are not attempting to only achieve the highest prediction accuracy on our data set (although we will attempt to do so) but to understand how our different approaches impact the end models. This purpose is intentionally non-specific such that our results can be generalized to other markets and used to better understand machine learning techniques for predicting investment returns in any context. One such emerging market where this research is likely to be useful is in the cryptocurrency markets. Cryptocurrency markets are structured in a fundamentally different way than traditional financial markets. This poses the question of whether the same machine learning modeling techniques that have been successfully used to predict market movements in traditional markets can also be used to predict market movements in cryptocurrency markets or whether new approaches are needed. Given that there has been few prior works investigating this problem, our research focuses on making some headway. We apply various different machine learning modeling techniques to data from both traditional markets and cryptocurrency markets. This allows us to analyze the results and determine whether or not the various techniques exhibit the same predictive tendencies in the two markets. Of course, any two markets (even if they are both traditional markets) differ in their structure and modeling techniques will therefore vary in their effectiveness. However, it has many times been exhibited that different kinds of machine learning modeling techniques perform best with certain kinds of data or in specific contexts. For example, it is widely known that the Long Short-Term Memory recurrent neural network model consistently outperforms other architectures when dealing with time series data in financial markets. It is an open question as to whether this holds for novel cryptocurrency markets. This is only one example of what our investigations may shed light on and we intend for this research to provide those that are interested in predicting rates of returns in any financial market with useful rules of thumb regarding when to utilize certain modeling techniques. Figure 1 below gives an overview our solution framework for determining the effectiveness of various machine learning techniques and algorithms for predicting an investment’s rate of return in both traditional markets and in cryptocurrency markets.
Beneficiaries of our project include trading firms, individuals, and researchers that have an interest in machine learning applied to financial data. By understanding the impacts that the type of machine learning model and techniques used have on our predictions for different assets in different market conditions they will be able to better decide which approaches they may want to use for their task. This is especially relevant in esoteric or emerging financial markets where asset price patterns are less established or correlated.