Good News: Using News Feeds with Genetic Programming to Predict Stock Prices
This paper introduces a new data set for use in the financial prediction domain, that of quantified News Sentiment. This data is automatically generated in real time from the Dow Jones network with news stories being classified as either Positive, Negativ
- PDF / 342,632 Bytes
- 12 Pages / 430 x 660 pts Page_size
- 60 Downloads / 184 Views
Abstract. This paper introduces a new data set for use in the financial prediction domain, that of quantified News Sentiment. This data is automatically generated in real time from the Dow Jones network with news stories being classified as either Positive, Negative or Neutral in relation to a particular market or sector of interest. We show that with careful consideration to fitness function and data representation, GP can be used effectively to find non-linear solutions for predicting large intraday price jumps on the S&P 500 up to an hour before they occur. The results show that GP was successfully able to predict stock price movement using these news alone, that is, without access to even current market price.
1
Introduction
Stock market price prediction has long been an attractive area for research, with techniques including everything from Neural Networks [1][3][2][4] to Genetic Programming [5][6] being used to try and predict stock price movement. These methods typically base their predictions on factors such as recent prices in the market. This is despite the Efficient Market Hypothesis (EMH) [11], which states that financial markets are “informationally efficient”, that is, stock prices immediately reflect all known pertinent information so that it is not possible to outperform the market using information which is already known to the market. While the EMH would write off any success to luck, effectively saying that one is as likely to have the same success rolling chicken bones as running GP, these predictive methods gamble on having access to high quality information that no one else has. In particular, although the same raw information (typically stock prices) is available to everyone, not everyone has the ability to analyse it in useful ways, and so, there is opportunity to profit while the market adjusts its prices, as an unused source of information may give investors an advantage. This paper considers a different source of information, news stories. Although the basic idea that there is a relationship between news events and stock market price movements is not a new one [7] there has been very little work done to M. O’Neill et al. (Eds.): EuroGP 2008, LNCS 4971, pp. 49–60, 2008. c Springer-Verlag Berlin Heidelberg 2008
50
F. Larkin and C. Ryan
incorporate news events into quantitative style models. This may be due to the fact that the human interpretive element of news stories does not easily lend itself to the quantitative scrutiny that is regularly applied to the so called hard data such as employment numbers or interest rates. However, if one could employ news stories/events in a quantitative and automatic way, then this could give one an enormous advantage in the market, in the sense that it would be possible to react more quickly than the market. Recently, a research company (RavenPack International, S.L.) has developed means for quantifying news stories; and our goal was to search for a model based on news sentiment (i.e. whether a news story is relevant to the particular market or sector, and if it
Data Loading...
