Using fundamental indicators and machine learning to predict stock returns

Abstract

Stock price prediction as a discipline has attracted a lot of attention through the entire history of the existence of stock markets. Successful price prediction can potentially result in huge profits. Proponents of the random walk hypothesis are of the view that stock price movements are random in nature and are thus unpredictable. However, many people have devoted considerable time and effort in trying to model price movements and today, with modern technological concepts and computation power, the output of such analysis has become that much more verifiable and useful. Stock price prediction is largely divided into two categories – fundamental analysis and technical analysis. In fundamental analysis, the past and expected future performance of a company is considered while valuing its stock. This involves taking a long term investment horizon and basing one’s predictions on expected industry and company performance among other things. A technical analyst, on the other hand, is concerned only with the historical price movements of the company’s stock. He makes use of various tools and techniques in trying to get trends out of the past price movements and basing his investment decisions on such shortterm predictions. With modern-day computers and advancements in the field of machine learning, newer methods of predicting stock prices are being used, either as standalone models or as an ensemble technique for combining different models. This is an emerging field in the area of stock return prediction. Our proposed work would involve the use of machine learning methods and artificial neural networks to combine various fundamental indicators for stock return prediction in order to add to the existing knowledge-base of this area of analysis. The proposed work primarily intends to predict stock return of companies listed on the Indian stock exchanges using fundamental information. The fundamental information that we plan to use includes key numbers and ratios from the annual reports of the companies. For instance, if we are analyzing a company in the pharmaceutical sector, then the R&D expenditure of the company can be a crucial indicator of the future performance and hence the future stock price of that company. However, for a steel company, a more pertinent indicator could well be the expenditure on new plant and equipment. Thus, relevant parameters will depend on sector-specific and company-specific details. The key indicators thus identified will be combined using methods of ML/ANN to give us an ensemble indicator of company’s future performance and stock prices. In addition to using standard API libraries for implementing the ML/ANN algorithms, we would also attempt to customize these algorithms in a manner suitable for the present requirement. The models will be trained on the training dataset portion of the population. Post training, we will test the models on the test dataset. A limitation of our study might surface in the form of fudging and misreporting of accounting information provided by corporate which will add noise to our indicators. This is an inherent limitation and expects to achieve a superior performance using our methods, despite the presence of such noise in the dataset. We collected data from Prowess database supplied by CMIE. We took our universal set as 30 companies’ which forms part of Sensex Index of BSE. The same was used considering that the companies are well researched and covered by analysts. This ensures that the financials of the company are reliable. Also since most of the trading volume in stock market is the stock of these companies, the market prices are more efficient than other mid sized and small sized firms. We took the entire data available for these companies in the database. We were able to retrieve quarterly historical financial information from September 1997 onwards. In total we received 1,578 data points for different companies and periods. We would be using this set to carry out various analyses and come up with our recommendation.