Market prediction based on macroeconomic indicators - page 11
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
Perhaps such naivety arises from the lack of implementation of this model?
Why haven't you implemented it yet?
I remember a couple of years ago there was a buzz about searching for progessors, during this time you could not just learn mql, you could become an ace.
It is impossible to transform an initial non-stationary series into an equivalent stationary series. It is possible to manipulate the original series in a variety of ways, but one must realise that the result may not be equivalent to the original series. This is what happens when one performs a "conversion of a non-stationary series to a stationary series".
Why is it impossible? It is very possible and has been done for decades. The proof is quite simple. If there is a model with non-stationary data:
(1) y = f(x1,x2,...)
then there must exist a model on transformed (differentiated) stationary data
(2) dy = df/dx1*dx1 + df/dx2*dx2 + ...
Where dy, dx1, dx2 ... are our stationary data. The transformation back to non-stationary data is quite simple:
y[i] = y[i-1] + dy
Finding models (1) of non-stationary data is quite difficult. Finding models (2) of stationary data is much easier. Let me also try to explain in a simpler way. If I give you the value of gross output for some quarter, domestically $100 billion (non-stationary input x1), can you predict the Dow Jones index (non-stationary output y)? No one in the world can solve such a problem. And if I tell you that gross output fell by 5% (dx1), can you predict the change in the Dow Jones index over the same period of time (dy)? This is much easier as it does not require knowing the absolute values of gross output and the index. At least the sign of the change in the index (minus) can be predicted with 100% accuracy. And we just need it much more to make money than knowing the absolute value of Dow, 15000 or 18000.
Although I don't deny that it is very important for rocketmen to know just the position of the target, not the increments of the target, in order to hit the target. Maybe that is why it is hard for missile engineers to make money in the market: they cannot let go of their notion of price as a moving target :)
Tried my model to predict GDP growth. It turned out pretty decent, the model found at least 3 predictors, each of which increased the accuracy of the predictions. The blue line below is the actual change in GDP, the red line is the predictions without looking into the future:
The model predicts GDP growth in the current (Q1 2015) and next (Q2 2015) quarter. The market should also go up.
Tried my model to predict GDP growth. It turned out pretty decent, the model found at least 3 predictors, each of which increased the accuracy of the predictions. The blue line below is the real GDP change, the red line is predictions without looking into the future:
The model predicts GDP growth in the current (Q1 2015) and next (Q2 2015) quarter. The market should also go up.
What is "predicting without looking ahead"? Is this section a forward model test?
Yes, forward test. Although even the forward test can be cheated if the predictors are selected with the knowledge of the whole history. For example, you may find a lot of articles recommending certain predictors of the market, such as gross product growth, unemployment rate, consumer price index, etc. And then predict the past based on those predictors without realizing that those predictors have been recommended based on all available history. In my case, the predictors and model are only chosen on data up to the quarter predicted.
Again, how is this graph constructed? Did you take a model built before 2000 and run it without retraining on this data or what?
Forward by how many values forward?
1)
Why is it impossible? It is very possible and has been done for decades. The proof is quite simple. If there is a model with non-stationary data:
(1) y = f(x1,x2,...)
then there must exist a model on transformed (differentiated) stationary data
(2) dy = df/dx1*dx1 + df/dx2*dx2 + ...
Where dy, dx1, dx2 ... are our stationary data. The transformation back to non-stationary data is quite simple:
y[i] = y[i-1] + dy
2)
Finding models (1) of non-stationary data is quite difficult. Finding models (2) of stationary data is much easier. Let me also try to explain in a simpler way. If I give you the value of gross output for some quarter, domestically $100 billion (non-stationary input x1), can you predict the Dow Jones index (non-stationary output y)? No one in the world can solve such a problem. And if I tell you that gross output fell by 5% (dx1), can you predict the change in the Dow Jones index over the same period of time (dy)? This is much easier as it does not require knowing the absolute values of gross output and the index. At least the sign of the change in the index (minus) can be predicted with 100% accuracy. And we just need it much more for making money than knowing the absolute value of dy, 15000 or 18000.
3)
Although I don't deny that it is very important for missile engineers to know just the position of the target, and not the increments of the target, in order to hit the target. Maybe that's why rocket scientists have a hard time making money in the market: they can't let go of their idea of price as a moving target :)
1) That's -- to put it mildly -- crap.
2) Demonstrate it with a concrete example. Thankfully, there are enough series in the terminal, and they are all non-stationary. What do you think the transformed series would look like? Stationary?
3) Rocket scientists know their stuff.
Think back to what they are
1) stationary random process
2) non-stationary random process
and what is the difference between them.
2) Demonstrate with a concrete example. Thankfully, there are plenty of series in the terminal, and they are all non-stationary. What do you think the transformed series would look like? Stationary?
Again, how is this graph constructed? Did you take a model built before 2000 and run it without retraining on this data or what?
Forward by how many values forward?
The example is in my first post. For the rest of your post, please speak out with at least some support for your conclusions, otherwise it looks like an argument with a 5th grader.
Well, I won't stir from my 5th grade cave... ;))
I will only note that this is your assertion from the first post
gpwr:
Without stationarity, no model will work.
is only true for the limited class of models that "your universities" have taught you.
But this limited class does not limit the set of all possible models at all.
It's clear from the first post that a lot of work has been done. And, I think, not unhelpful. Good luck.