stock market analysis markov models
TRANSCRIPT
Machine learning in stock market analysis
Agenda
• Economic concepts
• Can we predict the future price of a stock?
• Hidden Markov Models
• Building a virtual investor
• Experimental results
• Demo: Ben Investment Assistant
• Conclusions and future work
Economic concepts
• Stock Markets
• Stock price and volume
• Other indicators
Prediction of stock prices
• Random walk and the Efficient Market Hypothesis
• Dow Theory
• Conclusions
Hidden Markov Models
• Intuitive description
• Example:
Building a virtual investor
• He learns from historical financial data
• Based on what he learned he makes decisions (Buy/Sell/Hold)
• What data do we provide?
Preparing data
• We apply the EWMA financial technique to eliminate noise by smoothing the series.
• We consider for each the day the rate of growth by applying the natural logarithm for the daily return
• How do we make use of this data?
Computations
• Modeling observations: Multivariate Gaussian mixtures
• Re-estimations:
– What is the probability of being at state 2 at time 4?
– What is the probability of being at state 2 at time 4 at mixture 3?
– How do we re-estimate the model?
ComputationsForward procedure:
Backward procedure:
Computations
Computations
The algorithm
Experimental results
• Tests conducted for 14 randomly selected companies from different sectors: financial, utilities, technology, services and healthcare.
• We obtained to over 100% in revenues, and we suffered losses only when a company suffered a huge depreciation in its stock price.
• A few examples...
Goldman Sachs (NYSE:GS)
Above is the account evolution for investing in Goldman Sachs during June 07 – June 08 (After a year it generated a 53.3% revenue)
Above is the Goldman Sachs stock price evolution (June 07 – June 08)
Royal Gold (NYSE:RGLD)
Above is the account evolution for investing in Royal Gold (It generated a 50.3% revenue in 97 days)
Above is the Royal Gold stock price evolution for the testing period
An extreme case I (NYSE:MBI)
Above is the account evolution for investing in MBIA. The system does a good job at minimizing losses (only 26.2% loss)
Above is the MBIA stock price evolution for June 07 – June 08
An extreme case II (NYSE:MBI)
Using Auto-regression trees. A 74.2% loss
Above is the MBIA stock price evolution for June 07 – June 08
Demo: Investing in Google
• Ben Investment Assistant was done using:
• Windows Presentation Foundation, Sql Server, Analysis Services, ADOMD.NET, AMO, .NET 3.5, C# 3.0, Linq to SQL on Windows Vista Business.• 3-tier architecture, highly scalable
Conclusions
• Due to our results we can invalidate the assumption that past data has no use.
• Because the algorithm behaves like an investor we can have losses if the company suffers a severe depreciation of value.
Future work
• If we let Ben make decisions on a diversified portfolio we might almost be certain of a profitable outcome.
• We can expand the vector of observations to include more data (for example a news index calculated with text mining and Google search API)
Thank you!