What to feed to the input of the neural network? Your ideas... - page 42
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 read a book called "Masters of the Markets" by Bill Williams.
It's by Tom Williams. The book is almost 30 years old and many things have changed a lot since then. In particular, the share of trade automation has increased many times.
Well, you can get a tick chart. The whole point is in ticks. I made 60% per day using a type chart. Moreover, by types you can understand who invested how much money, if you take the most common size of ticks for $1000 (minimum lot). Then if there were 10 ticks of 23 points up, and then 60 ticks of 4 points down (the price went up and down to the same level), then most likely the smart money buys, and the stupid crowd sells, that is, the price will go up
I wouldn't say that. The smart money is smart enough not to let themselves be noticed. Most likely, they will make it so that the necessary position is gained not for 1 big noticeable tick, but for several average unnoticeable ticks. This is not difficult with the use of algo-trading and is done in fractions of a second or units of seconds. When the book was written, there were no such opportunities and when big boy gave an order on the phone for a large lot, it was visible. And you're not going to give 100 orders one after another over the phone, and even if you did, it would take quite a while.
What is a large/strong tick? Imho it is an application of a large client / bank for currency exchange (not speculative) operation. There is no reason for the bank to split it into small orders for the purpose of disguise.
1. Let's synchronise the clocks! We have a set of predictors - our task is to select the effective ones, i.e. those that improve our index - let it be accuracy. We make a selection, build a simple wooden model and evaluate the classification performance on a validation sample - so for each agent. And so on - we get the result - we send agents to explore new coordinates.
As a result, we have a set of binary variables - a switch - on/off.
How do we graph the waves here and at what point?
2. efficiency per number of iterations/wasted time, besides I described 3 different methods globally - interesting to do a full comparison between them.
1. At the moment of efficiency evaluation.
2. If interested, do it)))
1. At the time of performance evaluation.
2. If interested, do)))
Understandable all...
1. it is better when the global maximum of the system efficiency FF is single and stationary. It will look like a stable solid island of the surface among a sea of waves (well, or waves of the sea). Robust parameters imply performance with similar system performance on new data. If there are no such stable islands, it means at least two possibilities: either the system has no robust parameters at all, or the entire FF (or one or more metrics included in it) is chosen inappropriate for the process.
IMHO, the condition of uniqueness and stationarity of the FF maximum is impossible due to the fact that the market itself is by definition a non-stationary process subject to many unpredictable external influences (no detrending or taking derivatives will save us). The only thing we can use for successful optimisation (and subsequent forecasting) is the relative inertia of the market, but of course, provided that we consider the most liquid instruments with large volumes of trades and participants. Then we can find a sufficiently wide FF wave, which, although moving with time, still gives a FF value close to the extremum at the step between optimisations.
There is a (high) probability that no wide wave is found in the FF. I would not call such a FF inappropriate to the process and throw it out immediately, but try to add another layer of meta-optimisation/forecasting - on the sequence of FF wave surfaces on history (i.e. generalise/formalise the step-by-step transformation of waves and be able to synthesise the waveform for the next step). Ideally, this would be logically built into Walk-Forward Optimisation, but I haven't got round to it yet.
IMHO, the condition of uniqueness and stationarity of the FF maximum is impossible to fulfil because the market itself is by definition a non-stationary process, subject to many unpredictable external influences (no detrending or taking derivatives will save us). The only thing we can use for successful optimisation (and subsequent forecasting) is the relative inertia of the market, but of course, provided that we consider the most liquid instruments with large volumes of trades and participants. Then we can find a sufficiently wide FF wave, which, although moving with time, still gives a FF value close to the extremum at the step between optimisations.
There is a (high) probability that no wide wave is found in the FF. I would not call such a FF inappropriate to the process and throw it out immediately, but try to add another layer of meta-optimisation/forecasting - on the sequence of FF wave surfaces on history (i.e. generalise/formalise the step-by-step transformation of waves and be able to synthesise the waveform for the next step). Ideally, this would be logically built into Walk-Forward Optimisation, but I haven't got round to it yet.
Let's imagine that we already have a lot of predictors that we can't wait to feed to the input of the NS. But, the computer can be overloaded by the excess of incoming data - we don't have millions of dollars for super computers.
This problem is solved by dividing the data into parts.
What to do in this case, what optimisation algorithm will be ideal for the task of selecting the most useful predictors, what kind of FF can be invented?
Here you can apply any discrete AO whose goal is to select from millions of predictors a subset on which the best model will be trained.
In essence, the AO will simply select columns from a huge matrix of predictors.
FF is something that should be maximised e.g. model's akurasi or profit or whatever....
and will it be more efficient than standard methods from the economic point of view? This question does not cease to worry the poor :)))))) What are your thoughts?
Considerations are such, it is the only way, but here in general it is not about AO, but about how to organise competently the recording, storage and retrieval of data, AO here is a small and not the most important detail of a large mechanism and this layman will definitely not advise anything useful
The problem, as always, is an old one - the set was somewhere in the 50s.
For me, the result of the optimisation was always secondary, I always put the result of the forward in the first place.
The problem was that the best optimisation results were usually given by a failing forward, and some of the mid-ranking players in the optimisation, gave a forward that was pointing north.
I wondered how to bring to the first line of the back optimisation such a pass with a forward that will not be super duper, but at least slightly directed upwards.
It is clear that it is possible (not always) with the help of a custom optimisation criterion, which is now (or maybe has been for a long time) called FF.
How to find the right FF? I have not thought of anything better than the method of scientific poke, so I did it according to this time-tested method.
Andrey Dik has recently mentioned my article, so it is just about how to find the FF by the method of scientific poke.
The essence is very simple, optimise with forward on the maximum balance of the Expert Advisor, write a bunch of FFs, all sorts of different on what is enough imagination.
We run the script and look at the back and forward charts of our FFs.
If we find a decent graph, we run the next optimisation with this FF and with a high probability we get the correct forward on the top line of the back.
Most likely, like in the article, I again failed to express my thought clearly.
I apologise for being off-topic.
I came across a saying:
If you get shit on, you've made someone shit on you.
I'm talking to those who shit themselves. Get out of here, you stink of shit.
For me the result of the optimisation was always secondary, I put the forward result of that optimisation first.
The problem was that the best optimisation results usually came from a failing forward, while some of the average optimisers would produce a forward with a northward direction.
I wondered how to bring to the first line of the back optimisation such a pass with a forward that is not super duper, but at least slightly upwards.
It is clear that it is possible (not always) with the help of a custom optimisation criterion, which is now (or maybe has been for a long time) called FF.
How to find the right FF? I have not thought of anything better than the method of scientific poke, so I did it using this time-tested method.
Andrey Dik has recently mentioned my article, so it is exactly about how to find the FF by the scientific method.
The essence is very simple, optimise with forward for the maximum balance of the EA, write a bunch of FFs, all sorts of different on what is enough imagination.
We run the script and look at the back and forward graphs of our FFs.
If we find a decent chart, we run the next optimisation with this FF and with a high probability we get the correct forward on the top line of the back.
Most likely, like in the article, I again failed to express my thought clearly.
No, you did a great job of getting your point across.
Discrete AO.)) There is no such concept, all AOs are discrete, discreteness is set by the step of optimised parameters.
What kind of a wood pigeon do you have to be to think that AO and any MO methods can search for something without FF in the presence of such a huge number of articles on the forum and everywhere else about optimisation and machine learning....
Anything that is "searched", whether by conventional optimisation or by super-technological methods of MO, will be found exactly what is described in FF. The FF describes what is to be found, and everything else is just a search strategy, be it AO or machine learning methods.
"searched" - in quotes, because without FF nothing can be searched in principle.