Machine learning in trading: theory, models, practice and algo-trading - page 3728
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 to multiply with what without data input? Neurons have model weights (parameters), not data.
Ok. A database storing weights and offsets. Relate the word data to the name of the database (storage), not the type of information stored. In wooden data averaging models - i.e. you can relate to the data you trained on. In NS as previously stated weights and offsets are stored.
The definition is not mine. Saw this definition in the news a couple of years ago "neural network based databases". I agreed for myself. Especially for the main tool I use, i.e. trees/forests.
So is the challenge of identifying a row as a SB already successfully solved or not?
I have not solved such a problem for myself.
But I can guess.
The training data MO can learn well. Decision on new data. If there are 1000 rows, there may be 50/50 deviations. The more there are - the closer it will be to a random answer.
If after experiments none of the models can answer on new data better than 50/50, then it's either SB or you just couldn't find significant features and settings. 1 of these 2 reasons, but not unequivocally SB/not SB.But there are also factors of the degree of training, if undertrained - it will also be random or close to it on new data. You may have to test several models with different degrees of learning to determine the SB. If all of them give 50/50 on new data, then SB. If some of them will also answer correctly on new data, then there are patterns and they have been found.
So is the problem of identifying a row as SB already successfully solved or not?
My experiments have shown that it is easy. Just the task is that it would be difficult to distinguish - until I returned to this task.
Forum on trading, automated trading systems and testing trading strategies.
Will a good strategy work on randomly generated data?
Aleksey Vyazmikin, 2024.09.07 01:39 pm
Unfortunately, it turned out that the samples for model building were not quite correct for GBPUSD and Random_01 - the average values were taken more often than once an hour.
So I had to redo everything, I trained on 3 variants of sample combinations
I used 13 different subsets of predictors, as a result I got such Accuracy on the sample for them when creating the model with 0.
As can be seen from the table, random samples of bars are well distinguished by predictors. But EURUSD and GBPUSD cannot be separated as well - the best indicator is 0.68. I have highlighted in yellow the sets that interest me - let's see the significance of the first 25 predictors in these models.
For Test_CB_Setup_0_000000006 - We excluded everything except oscillators, except ATR and Volume + iVIDyA, iBWMFI, iChaikin
Surprisingly, oscillators still showed not bad results, which indicates, as I think, that there is a dependence in the development of time series sequences. Especially important were the indicators of standard deviation, Average Directional Movement Index by Welles Wilder, MACD andTriple Exponential Moving Averages Oscillator.
For Test_CB_Setup_0_000000010 - Only MA-based predictors
Unexpectedly, I managed to get a good result on predictors from moving averages. Especially the importance of predictors on TF H4 attracts attention.
Ideas and opinions are welcome on the results of the experiment.
My experiments have shown that it's easy. Just the task is that it is difficult to distinguish - until I returned to this task.
We have a custom character. We run a script on it, and it makes a random verdict. Is there something like that?
We have a custom character. We run a script on it and it makes a random verdict. There's something like that?
I'll try to answer without the mathematical nonsense.
A series behaves as a SB if its increments behave like a high-quality random number generator. There are many libraries testing the quality of such generators. A sufficient number of such libraries and tests in each of them unambiguously shows that the question is not simple and does not have an unambiguous solution. SB-ness can be violated in many different ways, in fact, in an infinite number of ways.
For example, there are NIST tests that would be suitable for a "coin" SB test.
I'll try to answer that without the maths stuff.
Is it possible to have a bid/ask series, on which the TS can make money, but which passes the tests as a SB?
It just seems that in any SB series it is possible to replace 1% of information in such a way that it will turn the series into a "earning" one. And this 1% will not affect the verdict that it is SB.
what does SB have to do with non-SB? Wake up.
when you have to categorise natural_price_order vs artificial.
both are SB (random walk, non-deterministic behaviour or super-complex unexplored laws). Both have similar features, and there are pseudo-laws in private samples.
Such a classifier is a wildly expensive thing, more expensive than all the grails presented on the forum. As well as a generator that can "cheat" such a classifier.
Is it possible to have a bid/ask row, on which TC can make money, but which passes as SB according to the tests?
It just seems that in any SB series it is possible to replace 1% of information in such a way that it will turn the series into "earnable". And this 1% will not affect the verdict that it is SB.
I guess you can't do without maths mumbo jumbo at all. You just have to figure out what a stat test is and how it works.
Stat tests always come with an idea of possible errors of both kinds (false negatives and false positives) and their probabilities (significance and power), so the answer is rather affirmative.
We have a custom character. We run a script on it and it makes a random verdict. There's something like that?
There's no off-the-shelf solution. But, theoretically, there are fewer consistent patterns in SB, and this can be evaluated with predictors. Both independently and relative to a set of other trading instruments, having identified common patterns in them at the beginning (through model training).
But in any case the answer will be about a time series of hypothesis testing type, not 100%.
Ok. A database that stores weights and offsets. Relate the word data to the name of the database (storage), not to the type of information stored. In wooden models data averaging - i.e. can be correlated with the data trained on. In NS, as I said earlier, weights and offsets are stored.
The definition is not mine. Saw this definition in the news a couple of years ago "neural network based databases". I agreed for myself. Especially for the main tool I use, i.e. trees/forests.