- Joined
- 2/11/14
- Messages
- 17
- Points
- 13
Trading is the field where also the most brilliant minds can fail miserably. Nevertheless, one can still enjoy and fun.
I'm interested in analysing datasets in order to identify factors that may contain predictive information for a trading purpose. My very first attempt consisted in taking several daily time series of stocks and using several technical indicators (RSI, MACD, MA, ...) I've trained Support Vector Machines. My experiment produced much worse than those claimed in literature (e.g. Tay and Cao (2001), Kim (2003)), despite the general setting was extremely similar.
I want to move on, but there are many options out there and I must start from somewhere. Also, any further attempt could require a lot of time to be mastered, in the sense of not falling in the common pitfalls. Also, I don't want to just do a blind attempt to improve the model, press the button and wait for the algorithms to run.
- Dataset: maybe it's too difficult to find alpha in financial time series if compared with other datasets (e.g financial statements, financial news, Google trend, Amazon or something more fancy.)
- Feature selection: given a dataset, create features.
- Data pre-processing: clustering and dimensionality reduction can remove noise.
- Models: meta-modelling (ie Bagging, Boosting and Stacking) can significantly improve the predictive performance.
- Neural nets: maybe mastering the state of art algorithms requires a lot of time that could be used in the previous topics instead, at least for the moment.
- Implementation in practice: this is a whole other story. I don't want to discuss it in this topic.
In the long run I strongly suspect that one have to master all these key areas to get a solid result (assuming that it's possible). My first concern is that I would avoid spending years only studying daily log returns and technical indicators when there could be "better" data to analyse.
I would be interested to listen your opinion about some of the points that I've touched in this thread. Also, If you have tried to do something similar, I would appreciate to listen your story.
I'm interested in analysing datasets in order to identify factors that may contain predictive information for a trading purpose. My very first attempt consisted in taking several daily time series of stocks and using several technical indicators (RSI, MACD, MA, ...) I've trained Support Vector Machines. My experiment produced much worse than those claimed in literature (e.g. Tay and Cao (2001), Kim (2003)), despite the general setting was extremely similar.
I want to move on, but there are many options out there and I must start from somewhere. Also, any further attempt could require a lot of time to be mastered, in the sense of not falling in the common pitfalls. Also, I don't want to just do a blind attempt to improve the model, press the button and wait for the algorithms to run.
- Dataset: maybe it's too difficult to find alpha in financial time series if compared with other datasets (e.g financial statements, financial news, Google trend, Amazon or something more fancy.)
- Feature selection: given a dataset, create features.
- Data pre-processing: clustering and dimensionality reduction can remove noise.
- Models: meta-modelling (ie Bagging, Boosting and Stacking) can significantly improve the predictive performance.
- Neural nets: maybe mastering the state of art algorithms requires a lot of time that could be used in the previous topics instead, at least for the moment.
- Implementation in practice: this is a whole other story. I don't want to discuss it in this topic.
In the long run I strongly suspect that one have to master all these key areas to get a solid result (assuming that it's possible). My first concern is that I would avoid spending years only studying daily log returns and technical indicators when there could be "better" data to analyse.
I would be interested to listen your opinion about some of the points that I've touched in this thread. Also, If you have tried to do something similar, I would appreciate to listen your story.
Last edited:
