WEEKLY DIGEST 2015, January 03 - 10 for Neural Networks in Trading: "We obtained a very significant success in pre-dicting stock price the next day"
12 January 2015, 21:11
0
353
WEEKLY DIGEST 2014, December 27 - 2015, January 03 for Neural Networks in Trading & Everywhere: Neural network over-fitting
"Overfitting is not only when test error increases with iterations. We say that there is overfitting when the performance on test set is much lower than the performance on train set (because the model fits too much to seen data, and do not generalize well)."
Research paper evaluating the efficacy of several machine learning approaches to predicting stock market indices.
Made by Rohan Patel
"We obtained a very significant success in pre-dicting stock price the next day based on a day’s twitter sentiment. Going forward, we would liketo extend this analysis in several ways."
"Overfitting is not only when test error increases with iterations. We say that there is overfitting when the performance on test set is much lower than the performance on train set (because the model fits too much to seen data, and do not generalize well)."
============
Artificial Intelligence Archives by Stuart Gordon Reid
Good blog made by quant at KPMG
============
Using Artificial Neural Networks & Twitter Sentiment Analysis to Predict Stock MovementResearch paper evaluating the efficacy of several machine learning approaches to predicting stock market indices.
Made by Rohan Patel
"We obtained a very significant success in pre-dicting stock price the next day based on a day’s twitter sentiment. Going forward, we would liketo extend this analysis in several ways."
A. Improving models
- improving sentiment accuracy by having aclearer identification of high value tweetsvs. non-informative tweets. For this, a larger dataset would be required. Using word-dependency based models of tweet model-ing instead of a bag of words representationalso has the potential to increase sentimentclassification accuracy. As an example of amore subtle model of a tweet, one can insteadrepresent a feature vector as a parse treewhich can be fed into an SVM as a piece of non-vectorial data. A custom kernel definingsimilarity between parse trees can then bedefined.
- improving stock price prediction and general-ization of our analysis by considering multiplemonths over multiple years.
- considering non-standard neural network topologies such as recurrent and convolutionalnetworks as potential models beyond the feed-forward methods used in this paper.
- adding a neutral category for tweets as wellas buying decisions. Currently, even a mildlypositive sentiment and tweet would lead to abuy decision, which may not be optimal in areal-world setting.
B. Translation to investor decisions
Of course, it is worth noting that the currentanalysis doesn’t translate directly into a buyingor a shorting decision. There are several costsassociated with converting our results into a real-world process:
- Significant costs are associated with creatinga portfolio and maintaining it throughout the years
- Frictional costs could limit the profit gener-ation, especially as our analysis recommendsdaily buying and selling decisions, which issmaller than the typical decision time (about 6 months buying for retail investors)
C. Translating model for institutional investors
It is also worth noting that our current analysisis for retail investors. The other type of investors- institutional investors - enjoy significant benefitsover retail investors, such as:
- Lower transactional costs
- Higher liquidity allowing for greater overalldollar-valued returns
- Access to after-hours tradingThese could translate to a higher accuracy if this model is translated into decisions for insti-tutional investors, after appropriate adjustments tothe model. For example, our model can accountfor one of these - access to after-hours trading -by using change in closing price from one dayto closing price the next, versus the current modelthat uses the next days’ opening and closing prices.
