Machine learning in trading: theory, models, practice and algo-trading - page 238
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
I have now made an adequate selection of model parameters, trained, predicted, now got 0.69102. With the last dataset it was 0.69121, this week it is better, but it is due to chance, the model is essentially the same. Next week it will throw me back to +-0.0002 again.
For me this is the limit so far, my model is trained on 21 initial predictors without any tricks. For forex, for example I download different indicators from the terminal, select their parameters, etc. I.e. I get thousands of indicators out of 4 predictors (ohlc), and then I sift them out, leaving only thirty or so, and train the model.
I have sent them another file with all predictions=0.5, the score = 0.69315, you can use it for comparison.Ideally, I should somehow make thousands of these 21 numerai precursors, and sift out unnecessary stuff. But indicators are working with ohlc rows, not with such evenly spaced, I need to think how to generate more new ones.
No....
I get it, then I'll stop cluttering up the thread
I'm going to have to think about how to generate more new ones from them.
I have now made an adequate selection of model parameters, trained, predicted, now got 0.69102. With the last dataset it was 0.69121...
and I've lost everything, it's worse, 0.69120 minimum on RF, I tried MLP but it's not working at all
SZS look here rub the posts on some kind of strange algorithm((
I don't know what to do with it, but I'll try it if you do.) And all end the same way. This is normal + and the forum has long been spam. But the experiment can be carried out.
The more that the positive aspects of it, too, there is. Just look at how the trees are built on the randoms and the BP. How will the model-ts
on different randoms. And so on and so forth. I used to put both mei and astroindyuks and tide data in different ports of the world, and the migratory paths of
of animals, etc. that a normal person wouldn't think of.) Write and use whatever you want, just don't jump to conclusions.
#Sgenerate random, at least 10K observations
write.csv2(x, file = "D:/1.csv", row.names = FALSE, quote = FALSE) #write to file 1.csv on disk D:/
-----------
What a window))) By the way, this chelenge though is useless, but there are interesting moments. For example, how the participants drop in calibration data,
which allows us to speculate about the preprocessing data used by the authors of the task...
Forrest has long been over, he was only in the first two posts, then I wrote that I began to select patterns "my way" those when described when looking for good clusters - they are patterns.
I attach the pattern.
This is a trained cohonen
you feed it the last 10 median price values
MD <- median price
MD <- scale(MD,T,T) # normalize like this
library(SOMbrero) # run the package with the cohonen
MD[is.na(MD)] <- 0 # replace possible NAs
pred <- predict(model,MD) predict the cluster
if the cluster is number 41 (pred==41)then it's a buy, stop take to taste...
try it, maybe it will work for you too
I'm thinking, but I can't think of one)) , Who has any thoughts....
we have two candle configurations
According to the picture it is clear that in fact the candlestick pattern is the same, the difference is in the volatility, how can we mathematically bring these two patterns to the same value?
You can normalize it, but the zero point shifts, and this is important ...
like in the picture.
Just to explain, imagine that we want to know the color of the current, previous and previous candlesticks what we do:
close-open , close[-1]-open[-1] , close[-2]-open[-2]
We get a chart like in the picture "1", everything is clear and straightforward, everything above zero is a white candle, everything below zero is a black candle.
Now let's solve the volatility problem (the one that appeared in the first chart). We have to normalize the graph in a certain range - see picture "2" , but after normalization we get a new zero axis that does not contain any information.
How to normalize the data so that "zero is fair" and the problem with volatility is removed? what are your thoughts?
How to normalize the data so that "zero was fair" and eliminate the problem with volatility? what are your thoughts?
This is how I did it.
You have a total of 8 points on the chart - O(0), H(0), L(0), C(0), O(1), H(1), L(1), C(1)
After that these points O(0), H(0), L(0), C(0), O(1), H(1), L(1), C(1) can be assigned serial values 1,2,3,4,5,6,7,8
OHLC - open, high, low, close
(0) and (1) - bar number
These points can be arranged in descending order of price on the chart, from the highest to the lowest - H(0), H(1), O(0), C(1), L(0), C(0), O(1), L(1)
And now H(0), H(1), O(0), C(1), O(0), L(0), O(1), L(1) can be turned into a vector (2, 6, 1, 8, 3, 4, 1, 7), and normalized into 0-1 if needed.
And it will turn out that both plots will have the same "pattern" (2, 6, 1, 8, 3, 4, 1, 7), which describes the order in which these points go down in the plot
Unfortunately, I have not been able to squeeze any benefits out of this. With only 2 candles there can be 40320 patterns. Three candles = (4*3)! = 479001600 patterns, etc. Realistically it will be less, because, for example, H is always greater than O, H, L in the same candle, and C is less. But the number of patterns is still huge.
That is, you can assign a vector of numbers to any configuration of candlesticks, and use it to recognize similar configurations in the future. But the number of possible configurations obtained this way is so huge that it is probably impossible to find some graphical pattern after which the price will always go upwards/downwards. In the "Bill Williams' Fractals" strategy, for example, a pattern consists of 5 candles, and only a couple of combinations out of all the billions possible are traded.
Here's an illustration for 3 candlesticks for example.
Thank you no need))) I just suggested the randoms from different machines ...
You do not understand the idea, I can generate noise and myself a few times and do not need to shove it on different machines)
The essence of the idea is this.
I am training my MO for reversals. In the history of quotes there are not so many reversals... The quotes move in waves (waves) and in fact if you think about it the number of variants for reversal is quite sure, they are different mutations of head and shoulders, double tops, triple tops, etc.And these figures appear not because they have some mythical influence on the market but because the very variants of the reversal are finite in the wave movement, either like this or like that, or nothing else, and if we make a chart of the market, the cumulative Random, or something like that with the wave structure, we see that the reversals happen with the same figures, the same heads and shoulders will be in the Random...
So, in that article that was thrown by D. Trader, the author said that it is possible to generate some similar sampling for the network, in addition to the sampling that we already have and by that we increase the knowledge base of the network and thus increase the accuracy of the network.
So, since there are not so many reversals in the market, I figured that I can get an infinite knowledge base of reversals through cumulative randomness...
You see, I'm not looking for any mythical power that can predict the market ))))) And by checking this theory I got the result, which I posted, as far as I'm concerned the result is optimistic.
How to normalize the data so that "zero was fair" and to eliminate the volatility problem?
I did it this way -
Thanks, I will try....
I did it this way...........
If I understand it correctly, the method is too crude...
Let's take the simplest pattern of one candle
We have three three variants
they all fit into one logical pattern
O<H , O<C , O>L
H>O , H>C , H>L
C>O , C>L , C<H
L<O , L<C , L<H
If I understand correctly, your method classifies all three candlesticks as one pattern, and this is not good
How to normalize the data so that the "zero was fair" and eliminate the problem with volatility?
how exactly did you calculate the difference?
for example the difference between high and low
a = high
b = clowes
(a*100)/b so ?
I didn't get anything with that method, the network recognizes the hell out of it, even confuses candlestick colors
If I understand correctly, the method is too crude.