Autonomous processing of financial data for stock price predictions is widely used
by individual investors and corporations. successful forecasting however remains,
to a great extent, an unsolved problem. The number of factors that impact stock
price formation is large, and far too big, for us to be able to factor them all in.
Besides that, there is a widespread theory of efficient markets, that claims that
forecasting prices based on publicly available data is not possible, since the price
already reflects all publicly known information. There is also a question of how
to determine the success of forecasting models.
In this master’s thesis, we use support vector machines to predict the daily
closing stock price of four American corporations. The models used for forecasting
are both classification and regression models, the input variables used are historical
price data and technical indicators derived from them. The forecasting ability
of the models is then tested and compared to the random walk model, which is
considered to be the optimal forecasting model in an efficient market. Given that
we operate with financial data, the models are also tested in a simulated trading
environment using their predictions in a simple trading strategy.
For the analysis, we first look at the error metric between the prices we predict
using the regression models and the actual prices. We show that in some cases
we produce a smaller error that the random walk model. We then analyze the
percent of correctly predicted price movement directions, where we compare all
the models and show that only one regression model, does not outperform the
random walk model. Finally, we compare the returns and risk-adjusted returns
using a trading strategy, by far the best results are shown using the predictions
of the regression models based on technical indicators. It is also concluded that
the minimization of transaction costs is needed for a profitable trading strategy.
|