Re: Time Series Prediction Competition
- Hey Colin,
Or anyone else who is interested. Does anyone have any recommendations
for papers on intraday financial time series analysis? I have quite a
library of papers if anyone want to trade notes. Yes I am very
familiar with Sidhant Dash's excellent paper on the subject.
Also as I mentioned a couple of weeks ago. I will be putting up my
time-series specific NEAT implementation (based on Mat Buckland's
code) on Sourceforge soon. I am just waiting to get my project
approved. It works on windows and linux. Is relatively easy to use for
TS prediction. Has a couple of nice features like using a uniform
random number generator.
Anyone interested in chatting about NEAT TS or more specifically NEAT
TS for intraday financial prediction. Please drop me a line.
--- In email@example.com, Colin Green <cgreen@...> wrote:
> Thanks Ken.
> My initial thought is to try using my waveform generator experiment to
> find an approximation of the underlying function that generates the
> competition data, as at first glance some of the datasets appear to be
> very cyclical. Under the 'Instructions' tab on the web page it says
> the submitted results will be compared to "established statistical
> forecasting methods" and then it lists the following techniques:
> * Naïve
> * Single, Linear, Seasonal & Dampened Trend Exponential Smoothing
> * ARIMA-Methods
> So does anyone know off-hand what types of data sets these techniques
> can perform [good] predictions for, and therefore what type of data
> source is likely being used for these competition data sets? E.g. is it
> more likely to be, say, environmental readings such as temperature,
> river flow rates, water oxygen levels, population levels, etc. where we
> can expect some cyclic behaviour (through the seasons)?
> The reason I ask is that my investigations into financial time series
> data have lead me to believe you really need additional information to
> predict stock and indices price series with any sort of success, e.g.
> Price/Earnings ratios, and even then such series will possess a large
> amount of unpredictability caused by random events. I'll give it a try
> anyway but I think the chances of getting anything like good results
> with my existing waveform generator experiment are much greater with
> something like cyclical environmental data rather than financial data
> that, although cyclical in the very long term (like in decade
> timescales), tends to be largely chaotic IMHO.