Article: Price forecasting with neural networks

 
Forecasting Prices with Neural Networks

Many traders talk about neural networks, but few know what they are and what they can do in reality. This article gives a little insight into the world of artificial intelligence, explains how to prepare the data for a network, and also gives an example of forecasting with the help of Matlab.

Author: Shashev Sergei
 

Good afternoon!

A little familiarity with neural networks. It started with BrainMaker package, followed by MathLab. The leitmotif of this thread is the capabilitiesof neural networks, I can recommend to get acquainted with Tuckens theorem:

If a time series is generated by a dynamical system, i.e. values D_0 is an arbitrary function of the state of such a system, there exists a depth of immersion d (roughly equal to the effective number of degrees of freedom of that dynamical system) that provides a unambiguous prediction of the next value of the time series.

Skeptics, who appeal to the impossibility to predict as the probability of the next tick direction is 50/50 (up or down), may note that if it were true, then the mathematical expectation would be 0, and consequently on the long timeframes we would see a "straight" line.

But we see trends where the mathematical expectation is not equal to 0.

But in reality we see prices oscillating near some function, i.e. the process is STOCKASTIC.

To sum it up, despite the fact that forecasting seems to be EXTRAPOLATION of data, neural networks are actually solving the problem of INTERPOLLATION, which essentially increases the reliability of the solution. The prediction of a time series is reduced to a typical neuroanalysis problem - the approximation of a function of many variables by a given set of examples - using a series immersion procedure in multidimensional space.

Regards,

kirillov.

 
On behalf of the sceptics, I would like to point out:

The market is not a dynamic system.
The market is an OPEN stochastic system.
OPEN means that it is affected by many external factors.
And these external factors are not only uncontrollable (unmeasurable),
but even their multitude is uncertain.

In addition, the system itself is not constant over time.
Its elements (parts) can change their behaviour arbitrarily,
may sometimes succumb to the collective effect and may not.
Their behaviour in the system is affected by changes in the seasons, weather, solar activity,
even the phases of the moon...

The main parts of this system are human beings.

Hence we can conclude that predicting the SIGNIFICANCE of a price,
not only is not INTERPOLATION of data, but it is also not EXTRAPOLATION of data
(extrapolation implies a dynamic system).

For stochastic systems, we can talk about predicting their
statistical properties - probabilities, distribution functions, expectations, etc.
But again, provided that they (FRs, expectations, ...) exist and are constant over time.
 
I smell at least 10 pages of another discussion of neural networks in the forex market. ;o)
 
The future price depends on its previous price movements, which means that the most likely price trend can be predicted. Predicting the absolute value is a poor exercise, because even in different DCs the prices are different. But within one brokerage company the network gets used to its quotes and thus it can forecast the absolute value for a small period.

But it is possible to predict the direction with much higher probability than the absolute value :)
 
Mak:
On behalf of skeptics I would like to point out:

The market is not a dynamic system.

I disagree, because a dynamical system is a system whose state changes over time according to fixed mathematical rules; the latter are usually given by equations which relate the future state of the system to the current state. Such a system is deterministic if these rules do not explicitly include an element of chance.

The weakness of this formulation is "fixed mathematical rules", but no one has yet proved otherwise, and the entire history of forecasting relies on them.

Regards, Kirillov.

 
Hello! Like many people here, I've been making grids for time series forecasting at one time and have come to the following conclusions:
- Using grids to forecast exchange rates and even the direction of exchange rates proves to be less effective than using simple classical methods of technical analysis. Predictions of relatively simple grids do not exceed 70-75%.
- To get a 75% prediction or better, you need to build complex self-learning structures on supercomputers and spend years developing this stuff, and there is no guarantee that it will work.
- Grids are useful for solving a specific, well-defined tactical problem that is difficult to describe by statistical or mathematical means. Classification meshes and pattern recognition networks can be very effectively applied to solve tactical problems. There are some developments in this area, but it is very time-consuming and there is not enough time to do it. Whom it is interesting write, we shall work together: favorit_box@inbox.ru

P.S. In file materials of conference on neural networks. Interesting for lyknobesis.
Files:
 
solandr:
I smell at least 10 pages of another discussion of neural networks in the Forex market. ;o)


And I think so ;-)

But the quality of discussion will be at a higher level ;-)

 
VBAG:

- The use of grids to predict exchange rates and even the direction of exchange rates is less effective than the use of simple classical technical analysis methods. Predictions of relatively simple grids do not exceed 70-75%.

On behalf of practitioners I would like to point out:

Forecasting the direction of the currency rate at 70-75% is from the realm of fantasy.

I have been making such predictions for a long time, working through a bookmaker who takes bets on appreciation/depreciation of a currency over a fixed period of time (intraday). At first the bookmaker commissions were so small, that strategies with only 52% of correct predictions gave profits. At first, I used a simple system based on tehanalysis, which gave me about 54-55% profits.
Then the bookmaker commissions increased and I had to improve the trading system. I took all indicators I was using and put them into a neural network. The winning percentage increased to 59-60%. So there are tasks in which neural networks rule, regardless of the opinions of sceptics!
 
Better:
VBAG:

- the use of grids to predict exchange rates and even the direction of exchange rates is less effective than the use of simple classical methods of technical analysis. Predictions of relatively simple grids do not exceed 70-75%.

On behalf of practitioners I would like to note:

Forecasting the direction of the currency rate by 70-75% is from the realm of fantasy.
Perhaps we are talking about different percentages, but that is not the point. Widely known MACD, OsMA, Regression Analysis, etc. make predictions no worse than quite sophisticated grids. And often even higher. And my main idea was if we want to get a qualitative leap in comparison with classical methods we should create complex self-training frameworks using МtLabe or SNNS (or better yet write our own) but not rely on nice wrapped programs like NeuroShellDayTrader (total nonsense).
If we want to improve MACD prediction quality by several percents, it would be better to create a grid in one evening using good old NeuroSell2 or BrainMaker, compile it in C code (simple set of transfer functions with coefficients) and implement it in an Expert Advisor. It works quite well. But it will not solve the problem of how to become a millionaire.
 
VBAG:
Better:
VBAG:

- the use of grids to predict exchange rates and even the direction of exchange rates is less effective than the use of simple classical technical analysis methods. Predictions of relatively simple grids do not exceed 70-75%.

On behalf of practitioners I would like to point out:

Forecasting the direction of a currency rate at 70-75% is from the realm of fantasy.
Perhaps we are talking about different percentages, but that is not the point. Widely known MACD, OsMA, Regression Analysis, etc. make predictions no worse than quite sophisticated grids. And often even higher. And my main idea was if we want to get a qualitative leap in prognosis comparing to classical methods we should create complex self-training frameworks using МtLabe or SNNS (or better yet write our own) but not rely on nice wrapped programs like NeuroShellDayTrader (total nonsense).
If we want to improve MACD prediction quality by several percents, it would be better to create a grid in one evening using good old NeuroSell2 or BrainMaker, compile it in C code (simple set of transfer functions with coefficients) and implement it in an Expert Advisor. It works quite well. But it will not solve the problem of becoming a millionaire.

If I have forecast accuracy about 65-70%, is it enough for making profit on Forex? Did you get such percentage with linear regression analysis? Or by technical analysis in general (not on separate intervals, but on representative data)?