Machine learning in trading: theory, models, practice and algo-trading - page 1004
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 used to buy another product from them, there were no problems. But here it started.
How do the vendors assess the situation - is everything going as planned?
It's all logical: the trick-or-treaters are bound to be punished. ALWAYS.
It's more of a problem that when you buy something you don't know how it works, but you trust it... The question of basic trust - hello from childhood(a good childhood) - opinion.
Please
The formula for calculating the error is shown in the title of the table. Let me explain on the last example nnet: 204/(204+458) = 30,8%, i.e. the model produced 662 units in total, 204 of which were false.
The results are almost the same on 12 currency pairs, i.e. model performance is almost independent from the model and currency pair.
This result is obtained due to careful work with predictors whose predictive power changes very little when running a window of 500 candles over a file of 5000 candlesticks. The changes are within 5%.
PS.
I cannot show you the tester yet - it is stuck in the application of the tester for files over 1000 bars.
And what do you have as a target? ZZ sign?
Driven by the fierce desire to revive this branch, and taking into account that forecasting is possible exclusively and only on stationary VR
(1. Kolmogorov A. N. Interpolation andextrapolationof stationaryrandomsequences
2.Wiener N.Extrapolation, interpolation and smoothing of stationary time series)
Question:
In fact, the value CLOSE[i]-OPEN[i] is nothing but the sum of the increments.
A sequence of such values should, in the limit, tend to a normal distribution.
Well, there is an opinion that the sequence of returnees (CLOSE[i]-OPEN[i])-(CLOSE[i-1]-OPEN[i-1]) is a stationary series.
Has anyone tried such a thing on the NS input and what were the results???
P.S. Max, Doc, Mishanya, Koldun, Alyosha... Who are you throwing this thread at? А?
1) It should not. For example, it can have many different marginal distributions.
2) Probably wrong. I've already given you a counterexample to the vertex\done. Non-stationarity is "not a bug, but a chip" that appears as a result of market makers getting rid of "excesses" by the bulk of traders.
What do you have as your target? The ZZ sign?
Increment
ZZ can make an awesome trending teacher, but I couldn't find any predictors for this teacher - all give an error of about 50%And Kolmogorov, in general, I see, paid special attention to B(k)=M[x(t)*x(t-to)]=M[(CLOSE[i]-OPEN[i])*(CLOSE[i-to]-OPEN[i-to])] and refused to predict anything unless this function had a quite definite form.
Maybe it makes sense to put certain conditions on the NS work?
Say, skipping unsteady pieces of BP, examining second returns or B(k), for example?
The ARIMA model has similar conditions.
You can train a model and even make a profit on a chart, but if certain conditions and requirements are not met - you cannot trade with this model anyway. The Dickey-Fuller stationarity test.
In GARCH the distribution of predicted returnees is also observed to be similar to the original data.
I think much of what you want to do is already implemented in this model.
Regarding neuronics - you can't just shove some time series into it, train it to maximal result and wait for profit. This will lead to "overfits"- the neuron simply memorizes its existing data, and is unable to work adequately on new data. We need to choose its training parameters and sometimes stop training and make cross-validations to make sure that the overfit has not yet occurred.
If everything is done correctly, the training will stop quite early, when R2 is a bit higher than zero. There will be stable profit on the equity chart both on the training and new data but spread exceeding a couple of points will turn everything negative. For more accuracy the programmer should either use deep nets and weeks for training or select various indicators to feed neuronics with time series.
And Kolmogorov, in general, I see, paid special attention to B(k)=M[x(t)*x(t-to)]=M[(CLOSE[i]-OPEN[i])*(CLOSE[i-to]-OPEN[i-to])] and refused to predict anything unless this function had a quite definite form.
Maybe it makes sense to put certain conditions on the NS work?
Say, skipping unsteady pieces of BP, exploring second returns or B(k), for example?
No, it doesn't work.
There's one good thing on the market for phicha engineering - the interconnection of fintechs. You can create similar but slightly different instruments and watch the distributions between them. Everybody is tired of deriving features from one BP :)
Increase
I can make an excellent trend teacher out of ZZ, but I was not able to find any predictors for this teacher - all give about 50% errorZZ is a tricky tool for targets, with it you need to prepare the sample in a special way, so as not to cheat yourself.
Accuracy "around 50" is quite normal, if you use ball data, above 53% you can trade, and in general accuracy for this is a shitty metric, it can easily be accuracy ~50%+-1% and correlation of predicted increments with market is >5% (0.05) and that is huge, not a grail of course, but enough to trade in a portfolio with other strategies. Use either correlation or R^2, or logloss if you get used to non-linearity
And everyone is tired of extracting features from one BP :)
After all, we are dealing with 2 streams:
1. a stream of events - time of appearance of a new quote (intervals between them)
2. the price series itself in this flow of events.
I will try to prove (or disprove) that randomly changing the time of reading of quotes, the autocorrelation function behaves differently in one and the same moving observation time window.
What's my point?
А! My point is that "thinning" of a stream of events plays a big role. Perhaps - a key role. My son, Alyosha, won't lie. But as long as it's all "in the process..."
After all, we are dealing with 2 streams:
1. a stream of events - the time of appearance of a new quote (intervals between them)
2. the price series itself in this flow of events.
I will try to prove (or disprove) that randomly changing the time of reading of quotes, the autocorrelation function behaves differently in one and the same moving observation time window.
What's my point?
А! My point is that "thinning" of a stream of events plays a big role. Perhaps - a key role. My son, Alyosha, won't lie. But as long as it's all "in progress..."
So he's already admitted that he's been leaking for years, then he played it back, and now nothing works for him.
If there was a real pattern to the BP transformation, it would still work. Tales about the markets being more efficient now, and "I used to do great things" do not work.
Actually there is no one to unmask, no matter how hard they resist the outcome is the same :) So Aleshenka is a bad proverb, a new one is needed.