Kohonen and Patterns - page 2
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
Well, what I see on M15 is basically consistent with my understanding of proper NS operation. Indeed, as the forecasting horizon increases, the forecasting accuracy should decrease. Hence, we can conclude about the advisability of predicting only one step ahead (as the most extreme option), for example, daily bars. It would be interesting to look at results of such forecasting. Or, as an alternative, H4.
In the penultimate image despite the fact that the number of bars forecasted ahead is the same, it is equal to the length of day, but for training different steps of averaging 24,48,72.96 and 120 hours are used. That's why there are five forecast variants, though only the 24 hour one - it's red - is very different from the others.
Well, what I see on M15 is basically consistent with my understanding of proper NS operation. Indeed, as the forecasting horizon increases, forecasting accuracy should decrease. Hence, we can conclude about the advisability of predicting only one step ahead (as the most extreme option), for example, daily bars. It would be interesting to look at results of such forecasting. Or, alternatively, H4.
>> Then it is clear.One step ahead is not always good. This network is like a regression. While the regression is moving in the middle of the spread - everything seems fine, but when a strong deviation occurs the regression reacts and the end moves away from the spread. So if the alignment of the point from which the forecast is coming is good, then the forecast is OK. But if you follow the data all the time and predict only one step ahead, the number of errors can increase dramatically.
-
Hmmm... how complicated the world is!
Thank you. I'll have to think about it...
A 3 month forward forecast for the pound on the Daily, for the duration of the Championship.
-
A 3 month forward forecast for the pound on the Daily, for the duration of the Championship.
-
Man, if you can predict the market like that already, I envy you.
Man, if you can predict the market like that, I envy you.
Well, if you have looked at the previous page, you have seen that the net often lies. It shows more or less the nature of a possible trajectory, but the values themselves are often very different. For manual trading it is more or less the same where there is a continuous monitoring. However, it is difficult to make an Expert Advisor based on it. The readings are too ambiguous. Reliability checking is always needed. In addition, it strongly depends on the data being submitted for input. In the given prediction the data was that drawn in yellow and if you look carefully, the network picked out an area that is approximately slightly to the left of the middle and repeated it with some multiplication. And if you feed in more time data, the picture can be quite different. So not everything is as great as it seems so far.
It turns out, it is. And I'll delete this thread too if they mess it up again.
Is it allowed to delete threads here? :-OHOW?!!!
I looked at mine... I didn't see any controls for deleting a topic. You can't delete your own posts here after a while, but you can't delete threads...
Is it allowed to delete threads here? :-OHOW?!!!
Looked at mine... And didn't see any controls to delete the thread. You can't delete your own posts here after a while, and you threads...
Not anymore. It's been three days. So the thread will live on unless the moderators delete it, which is unlikely.
Forecast for 3 months ahead for the pound on the Daily, for the duration of the Championship.
-
What's on the inputs may I ask?
It won't work any more. It's been three days. So the thread will live on unless the moderators delete it, which is unlikely.
What's on the inputs, may I ask?
The inputs are the data from the previous year. They are shown in yellow on the chart. This is Close pre-filtered a bit. So we have a period of known data T of approximately 260 days. We have a task to extrapolate the forecast for 3 months ahead, or about 66 days. We denote the extrapolation period by Te. Now let's input the known data for the period Ti or input and the data for training Te or output in one day increments. Then we run through the entire year. After training, we read the last known data for the period Ti and make a forecast. In blue we draw to check how the network has learned. In blue before last data is how the network can predict 1 bar ahead and in light blue already beyond the known data is the forecast.
The inputs are the data for the previous year. These are shown in yellow on the chart. This is Close pre-filtered slightly. So we have a period of known data T of approximately 260 days. Our task is to extrapolate the forecast for 3 months ahead, or approximately 66 days. We denote the extrapolation period by Te. Now let's input the known data for the period Ti or input and the data for training Te or output in one day increments. Then we run through the entire year. After training, we read the last known data for the period Ti and make a forecast. In blue we draw to check how the network has learned. Blue is before last known data, it is how the network can make 1-war forecasts and light blue is outside known data, it is the forecast.
As far as I understand it -- sliding window learning?
Hmm, how do you look at the rather simple (by code, not by result) assessment of the validity of the result?