Econometrics: one step ahead forecast - page 39
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
What do we care about the size of the error? It may be 1000 points. You can regulate its use and size yourself - you can enter not by the bar opening, but by a pending order near the error line.
You will have a different forecast before the opening of the bar, and you will always be out of fluctuations of the price.
But if you want to use this method, you should make a forecast for at least 500-1000 steps to estimate the confidence interval boundary.
Once again. There is a concept of predictability. God forbid one step forward. everything else is bogus. At 500 steps the error will be 500 times larger. There is a concept of "dynamic predictability". Its graph is such a rapidly expanding funnel.
The concept of prediction error is basic. There are plenty of forecasters on this forum who can't answer the question about forecast error. Although it is obvious that if the forecast + error is greater than the candlestick length - then the trade is not possible.
Пока дождетесь, будет другой прогноз и всегда будете вне колебаний котира.
This is not true - you yourself say that the deviations are large by results - i.e. the order will often be captured by the price.
Once again. There is a concept of predictability. God forbid one step forward. everything else is bogus. At 500 steps, the error will be 500 times greater. There is a concept of "dynamic predictability". Its graph is such a rapidly expanding funnel.
If your method cannot predict at every step - wait for the right situation. Not by 500, but by the root of the square of the current error and the square of the error in the previous step. Therefore - this funnel expands more and more slowly with time. That's why I told you about 500 - 1000.
Although, it is obvious that if the forecast + error is greater than the candle's length - then trading is not possible.
This is not the case - you yourself say that deviations are large in terms of results - i.e. the order will often be captured by the price.
If your method cannot predict at every step - wait for the right situation. Not by 500, but by the root of the square of the current error and the square of the error at the previous step. Therefore - this funnel expands more and more slowly with time. That's why I told you about 500 - 1000.
No, I have already explained why.You know best. Everything I write about has been written by many people before me, your thoughts are big news to me.
It's just that I told you about the option of trading on a rebound from probability boundaries - as in price-chenel. So as not to be afraid of making too much of an error. You wanted to try different models. But the estimation of the boundary should be more or less stable for this, a lot of points are needed.
You can make a forecast on D1, get a forecast on 1440 minutes. But the forecast error will be very high for minutiae. I think so.
One thing has now occurred to me. I haven't considered using the forecast until now. So maybe you are right about the reversal or rebound strategies from the boundary.
If you look at the forecast tables above, the error is at least 50 pips. But it can be reduced.
I like your thoughts more and more as a variant of using the forecast.
faa1947, I think we should start by looking at an elementary market model.
Here is model 1 for example. There are only 2 participants trading an asset. At the initial moment of time they have equal amount of money, for example X=1000 and equal amount of this asset (let's call it a stock), for example Y=100. What property will the price change process have? Obviously, the maximum price of the asset will be limited by the money in hand. There will also be a level of "fair price" - there are only 200 shares, and 2000 money in hand. Fair price=2000/200=10. At this price the turnover of shares will be maximal. If the price is lower, not all of the money is used, if the price is higher, not all of the shares are used. It is logical that to maximize the turnover the price should fluctuate around this level, because a constant price is not interesting for the traders.
If more money comes to the market, the fair price will be higher. Accordingly, a real trend will occur. It is very difficult to separate it from the fluctuations around the fair price. The larger is the value of fluctuations relative to the trend component. But the econometric model will be like yours - a certain fair price that changes under the effect of inflow/outflow of funds into the market and the noise fluctuations around it. It is also kind of clear how to make profit - to trade back to the fair price at extreme deviations.
Let's complicate the model №1 - add participants and it will be model №2. And some of them will be intraday traders (open and close positions within 24 hours), while others will hold positions for several days (as long as they want). One fair price is left, but the intradayers also create some price fluctuations. That is, the fluctuations have become two-level - intraday and long-term. The fluctuation range is determined mainly by the amount of money and shares the players have at the corresponding level and their willingness to take a risk (risk appetite). Intraday speculators do not create a trend, only fluctuations. The econometric model will be more complicated - there are two noise components with different distribution parameters. But it is possible to remove the intraday noise component - consider diaries, which you have done.
Model №3 - there are many trading horizons and they do not have distinct beginning and end as in the case of the intraday model №2. Now there are as many noise components as there are trading horizons. Purely formally, we can introduce one noise component as a sum of all of them. But it only in case we trade only the maximum horizon. If we trade one of the intermediate horizons, the fluctuations of the higher trading horizons will not be noise for our level, but will be a trend, while averaging/smoothing should eliminate fluctuations of the lower trading horizons. What matters is the averaging period - too big and we lag behind the trend component, too small and we take the oscillating component as a trend. And should this period be constant? Of course, if those oscillations had a fixed period, like a sine, there would be no question about the averaging period. But there is no reason to think so - they are not periodic. Therefore, the averaging period should not be constant, or another version of smoothing should be sought.
Generally speaking, your model can be extended in the following way: trend component is a return to some average or smoothed average of a larger period, while noise is the price deviation from the smoothed one of a smaller period.
faa1947:
I like your thoughts more and more as an option to use the forecast.
faa1947, I think you need to start by looking at the elementary market model.
I have nothing to say about your model.
I like the (quality word) models based on crowd psychology best. In this thread C-4 gave a link to such a book. But that's on a "like" level.
The model I use is constructive: kotir = trend + seasonality + cyclicality + noise + outliers.
What is constructive is that I take trend + noise from the formula. There is no seasonality in forex. Cyclicality is very interesting, but I don't have any approach to this problem. I ignore outliers (averaging). This is constructive. If you include something in a formula, you need to understand a) what it is and b) how to model it. At the moment the "noise" is not fully considered on this topic, and the trend could use some work as well.
purely in pictures it is not badly illustrated http://kroufr.ru/content/view/3606/81/
Opinion is negative. another approach in TA. We take a history, calculate something from it and make a prediction.
It's all in the residuals from the model. If these residuals are stationary (mo and variance = constants), then the forecast is possible, if not, then it is not possible.
Here is a graph of the standard prediction error:
What will be the error in predicting 1 step ahead? Because you have to predict not only the trend but also the error. The whole problem is the error.