Hi @all!Does anyone have some experience with Singular Spectrum Analysis (SSA) in Trading? My results of a calculation are looking quite nice...

in sample or out of sample ?

Is there any difference ? What I am interested in is whether there is a opportunity to trade successfully with Singular Spectrum Analysis and whether there are some practical things which I have to attend.

Yes, I've been doing quite a lot of SSA of late.For those who haven't seen it before, the technique involves performing a PCA over lagged values of a time series. Each eigenvalue/eigenvector pair will represent a trend or cyclic component, from most significant to least. The nice thing is that all of the constituents signals can be added back to form the original time series.As acastaldo suggests, be sure that you form your Toeplitz matrix on in sample data only.If you're in London, drop me a line and we'll meet up over a coffee to compare notes. Or if you have any interesting ideas/questions/findings please post to this thread. Stewart Buttonwww.OnyxFinancial.co.uk

take care of the margin areas ... SSA/Fourier is neighbourhood dependent.I once worked in a fund incubator. A PhD (!) in Physics (!!), lecturing at a university chair (!!!) came in. He has the best filter - a butterworth, it fits the curve perfectly he said. He showed us graphs where we saw the entire chart and the butterworth output - "just look how perfectly it forecasts (!!)". Once I took his lovely butterworth and applied in a walk-forward manner he wasn't seen anymore.

A very basic Excel and VBA implementation available Here. The issue about neighbourhood dependence is well made. The non-causality is caused by the diagonal averaging or re-hankelisation in the final reconstruction step which causes re-drawing when new data arrives or the window period is expanded. Can be made causal fairly easily by only recording the end point or forecast values and iterating over some window period although this is obviously more computationally expensive. Also depends on the degree to which there is any underlying structure to discover in the first place.

Thanks for posting the spreadsheet. There are some good references there, especially on forecasting. Nice one!

Hi @ all,First: Thanks purbani for the wonderful spreadsheet especially the references!Second: My Problem with SSA and with i.e. Kalman Filter is the following: If you do this kind of calculation with a time series, let´s say of 100 data points, you will get a fabulous and nice filter result, but if you remove the last data point of you time series (data point 100), the calculated solution of either Kalman filter or SSA for point 99 will be different to the solution before (with 100 data points). So in my view the original and of course right calculation is curve fitting and this is why past results will be locking fantastic and future result will look badly? Finally, in my view, the calculation methods has to be changed: calculation for 99 data points with the only result for the 99th data point has to be totally delimited from the calculation for 100 data points with the only 100th solution? What are your opinions to that?

Looks like 20/20 hindsight. A lot of statistical models are like a photograph and describe the scene in front of you but tell you nothing about what is happening next. Beyond this it comes down to best practices in testing to ensure they capture what you want not what you have.

I'm trying to implement a trading system on SSA to se if it could be a valid tool and I would appreciate if you could share what are you experimenting. Let me briefly describe you what I'm trying to do. 1. I compute the SSA on a stock. 2. Analyzing the eigenvalues I try to separate the trend frequencies from the cycles (which is already a tricky problem...) in order to understand that is the trend3. I execute an SSA forecast of the trend and a SSA forecast of the filtered series. The prediction is made using the recursive method described by Tsionis rather then the alternate LLR approximation because I did not understand that yet...:-)4. Once recombined the series + the prediction, I filter the series by Savistkzy-Golay digitail filtering and compute the derivative first and second for both the trend +forecast and serie + forecast.5. At this point the signals are generated by: BUY if the derivative first is 0 (turning point in the series) and if the derivative second (the accelleration is an high negative value...so big accelleration turning up) . SELL if the opposite.6. In positive trend only long trades are considered and shorts if trend is negative...For testing I have my own backtesting test based on portfolio simulated trading.Does it make any sense what I'm doing ? Do you have some other ideas that could help me out ?regardsAntonio Signire