|
FAO: Krypton and other financial gurus
|
View this Thread in Original format
| Meat187 |
Just a quick request:
I want to demonstrate some data analysis techniques on stock charts. For this I need the charts in a table that Excel (or preferably MatLab) can handle. I was thinking about DAX 30 over 1 year (would be 30 curves with 365 data points) but basically it doesn't matter which stocks I use.
My problem is where to get this data. Is there any place where I can just download some file, table, whatever giving me the values for a certain stock?
Thanks! |
|
|
| Moongoose |
| Google Finance or Yahoo finance, take your pick. Youll just need to do a few minutes worth of sorting the raw data because the export data option has the annoying habit of putting all the data (date, open close, hi, low...) into a single column. |
|
|
| Meat187 |
Thanks, but I can't seem to figure out how exactly to get the data.
So I went to google finance, looked for some company stock and then got this nice interactive chart. There I select 1 year and now where is the export / download button? :conf: |
|
|
| Meat187 |
Never mind, I found it.
Now is there a convenient way to do this for all stocks in a certain index at once? |
|
|
| Moongoose |
Hope that clears it up. |
|
|
| Acton |
| Out of curiosity, what type of analysis do you plan to work on? |
|
|
| Meat187 |
| quote: | Originally posted by Acton
Out of curiosity, what type of analysis do you plan to work on? |
It's not really work, more of a demonstration.
I got asked to do a 2 part lecture on multispectral imaging and unmixing. One of the unmixing techniques I want to present is PCA and my idea is to show that this concept is not only applicable to image analysis but that the "unmixing problem" is a more general one. The idea with stock analysis is to show that PCA can be employed to decorrelate / unmix a stock portfolio in order to minimize the risk. |
|
|
| tachobg |
| quote: | Originally posted by Meat187
It's not really work, more of a demonstration.
I got asked to do a 2 part lecture on multispectral imaging and unmixing. One of the unmixing techniques I want to present is PCA and my idea is to show that this concept is not only applicable to image analysis but that the "unmixing problem" is a more general one. The idea with stock analysis is to show that PCA can be employed to decorrelate / unmix a stock portfolio in order to minimize the risk. |
Sounds pretty neat. I'm curious how PCA leads to minimizing risk. Is each datapoint just the time series of prices for a particular stock, or? Also, is PCA generally a good enough model for stock prices (i.e., how well do the gaussianity and linearity assumptions of PCA hold?) |
|
|
| Acton |
| quote: | Originally posted by Meat187
It's not really work, more of a demonstration.
I got asked to do a 2 part lecture on multispectral imaging and unmixing. One of the unmixing techniques I want to present is PCA and my idea is to show that this concept is not only applicable to image analysis but that the "unmixing problem" is a more general one. The idea with stock analysis is to show that PCA can be employed to decorrelate / unmix a stock portfolio in order to minimize the risk. |
Interesting. But doesn't PCA work with clusters of data, whilst your stock portfolios will be set points of one value, per measure of time? |
|
|
| Meat187 |
I don't even know how well established this is in financial analysis (maybe someone who knows can elaborate / correct me) but my idea was this:
We take a look at the DAX 30 stocks and want to optimize a portfolie consisting of those. The data are the 365 values over the last year, so that forms a 30x365 matrix. Now some of those stock are obviously correlated, like BMW, Volkswagen and Daimler. Buying those three would not minimize the risk much in comparison to buying just one of them. The goal is to find an uncorrelated selection of stock in appropriate quantities, so that they won't affect each other and therefore hedge the risk in a better way. PCA can do exactly that, it finds principal components that represent as much data variance as possible while being uncorrelated among each other. Then the portfolie can be selected to match the first principal components.
A very simplistic approach, of course, and again the main goal is to show how PCA is used for unmixing and that similar problems also occur in completely different fields than imaging. |
|
|
| Acton |
| quote: | Originally posted by Meat187
I don't even know how well established this is in financial analysis (maybe someone who knows can elaborate / correct me) but my idea was this:
We take a look at the DAX 30 stocks and want to optimize a portfolie consisting of those. The data are the 365 values over the last year, so that forms a 30x365 matrix. Now some of those stock are obviously correlated, like BMW, Volkswagen and Daimler. Buying those three would not minimize the risk much in comparison to buying just one of them. The goal is to find an uncorrelated selection of stock in appropriate quantities, so that they won't affect each other and therefore hedge the risk in a better way. PCA can do exactly that, it finds principal components that represent as much data variance as possible while being uncorrelated among each other. Then the portfolie can be selected to match the first principal components.
A very simplistic approach, of course, and again the main goal is to show how PCA is used for unmixing and that similar problems also occur in completely different fields than imaging. |
I see what you mean, my mistake. I thought you were going to take one time series and do an analysis on that using PCA, lol. But yeah, multiple DAX series could work, it would be really interesting to see what the analysis churns out actually, I've not used this method on financial markets before.
Just make sure you do an analysis over '08 for example, so you have some actual data to compare your results to. Then post them on here ;) |
|
|
| Meat187 |
| quote: | Originally posted by Acton
Just make sure you do an analysis over '08 for example, so you have some actual data to compare your results to. Then post them on here ;) |
What would I compare them to? |
|
|
|
|
