Econometrics: one step ahead forecast - page 119
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
This is just a commentary on the numerous points of the topic and the disagreement with them is this:
Trend is the increment on a set of values by a certain lag, to previous lags and in such lags there may be more than one. How to calculate this increment and assume a dependent increment for the next lag is a prediction model. At the same time methods of determination of significance of variables, just use step forward as the criterion, but not at all on lag - I wonder why with such common practice someone suddenly expects to get any guarantees of accuracy of forecast of exactly trend. Friendship with such "medical fact" is a straight carpet to a specialist psychotherapist... It goes without saying that the error accumulation will grow together with the lag size, but it does not mean reduction of forecast accuracy - for this measure is relative and is set by the correlation quality estimation, not the error size... Therefore, the choice of a model and its parameters is only a secondary problem, solved (and easily) after determining the size and properties of the sample of dependent variables...
Of course, one cannot use the word "error" without the words relative or absolute.
I did models on levels and on increments. If you compare correctly, the prediction error on the increment model was less than on the levels. When trading of course we are interested in increment and increment prediction error.
Unfortunately, the accuracy of the prediction has not affected predictability one step forward. I think one should predict the probability of direction one step forward rather than seek to reduce the forecasting error.
Perhaps it is possible to do it that way. I have decided differently, but I would like to hear other suggestions - modestly silent about mine (I assume that I have the moral right to collect a dividend for it in this form)... What confuses me about RMS is that it is equally "purple" for deviation in either direction from the mean, except that the regression will also turn out to be linear, for example - no one has promised either...
The RMS is fine if it is at least nearly constant. But if there is heteroscedasticity, it is definitely not. And heteroscedasticity can be different.
I had an idea to take the window size proportional to the number of bars in ACF up to zero correlation - it is the current memory length of the market.
I don't get it again. What evidence? It's a picture online! The lines are not being rearranged. Already said several times. It's not an FFT!!! It's a filter! How it is made will not be revealed.
This is the shift!!!
Each new bar from left to right = new picture.
Take the widest pitch of the lilac sinusoid
The amplitude changes and the periodicity changes. Right or wrong?
Amplitude changes and periodicity changes. Is it true or not?
This was covered in our conversation. There are four problems for prediction. The higher the frequency, the less distorted the period and amplitude.
The first two problems are solved by increasing the conversion window. the others are harder to solve.
This was covered in our conversation. There are four problems for prediction. The higher the frequency, the less distorted the period and amplitude.
The first two problems are solved by increasing the conversion window. the others are harder to solve.
I told you. The bass frequencies are more distorted than the upper frequencies. They are crooked. To get rid of this you need to increase the conversion window. Then these distortions will be out of sight. You'll be left with almost straight lines. But this transformation will also take more resources and time.
So it is difficult to extrapolate a curve with amplitude and frequency modulation. But it is easy to extrapolate an almost straight line. Of course, the problem does not disappear. Just take a small part of this curve that we can extrapolate with the smallest error.
Got it?
Decided after all. One bar can account for non-stationarity with a high enough accuracy.