Dear All,
I am using ANN with Wavelet transform for time series forecasting in some price data. ANN results are very good for fitting past data but when applied for predicting future price it shows a curve having exponential decay nature. Why ANN is giving that kind of results? Am I missing something.......?
Suggestions regarding this issue would be helpful.
Sincerely
Well as John Maynard Keynes said, "We all see the future with blind eyes" or Yogi Berra, who said, 'It's tough to make predictions, especially about the future'” . What you have observed is the farther you are from your data the less reliable is your prediction. Find a good reference on prediction by ANN esp. ones you are interested in and see what is going on in your field. Best wishes, D. Booth
Hi Devnath Shah
you could remove the trend of the time series. Then, you could use ANN to predict the de-trended time series. Simultaneously, you could forecast the trend with ets or ARIMA. Finally, you combine the trend and de-trended time series.
Best,
André
Dear A. Bauer Sir,
Thank You for your kind response.
In one of my recent work on hourly electricity market price forecasting, I had extracted the low-frequency component and high-frequency components using wavelet decomposition (1 level wavelet decomposition). Then ANN is used for fitting and forecasting the low frequency component. But for high-frequency fitting and forecasting, ANN gives a substantial error. One of my colleagues suggests to used GARCH method for modeling high-frequency component. How GARCH method can be used for forecasting high-frequency component. The nature of the high-frequency component is similar to that of white noise. Any tips or suggestions or MATLAB programs in this regard would be very helpful.
Sincerely
Obviously, the specifics matter so take what I say with a grain of salt but two things: 1) it sounds like you're dealing with the very popular "bias-variance tradeoff" problem. google it for more details. 2) It's probably better to try to predict returns rather than prices. both are generally non-startionariy but returns are less non-stationary compared to prices.
may be there are any problem at your rules, or intervals, or maybe your linguistic.
machine learning methods haven't traditionally done well with time series forecasting. Just consider the M-competitions and the GEFCOM competitions.
Traditional time series approaches do quite well, ARIMA, Double seasonal Holt-winter-Taylors and linear regression techniques as well.
GARCH is a common approach as it models the variation. See for example, price forecasting papers by Florian Ziel and Rafal Weron for accurate price forecasting models.
it's generaslly not a good idea to try to predict prices because they have a lot of autocorrelation in them ( price is essentally a cumulative random walk and random walks cause problems because of spurious regression. see the seminal paper by granger and newbold ) . you're better off trying to predict returns. You can do this by just defining the return as the difference of the log prices. Note that , by differencing, you're removing the random walk-ness feature in prices.
- Badges
- Science topic
- Similar topics
- Filing