New article Forecasting Financial Time-Series has been published:
Forecasting financial time-series is a required element of any investing activity. The concept of investing itself - put up money now to gain profits in future - is based on the concept of predicting the future. Therefore, forecasting financial time-series underlies the activities of the whole investing industry - all organized exchanges and other securities trading systems.
This article deals with one of the most popular practical
applications of neural networks, the forecasting of market time-series.
In this field, forecasting is most closely related to profitableness and
can be considered as one of business activities.
Forecasting financial time-series is a required element of any investing activity. The concept of investing itself
- put up money now to gain profits in future - is based on the concept
of predicting the future. Therefore, forecasting financial time-series
underlies the activities of the whole investing industry - all organized
exchanges and other securities trading systems.
Let's give some figures that illustrate the scale of this forecasting
industry (Sharp, 1997). The daily turnover of the US stock market
exceeds 10 bln US dollars. The Depositary Trust Company in the USA,
where securities at the amount of 11 trillion US dollars (of the entire
volume of 18 trillion US dollars) are registered, marks approximately
250 bln US dollars daily. Trading world FOREX is even more active. Its
daily returns exceed 1000 bln US dollars. It is approximately 1/50 of
the global aggregated capital.
99% of all transactions are known to be speculative, i.e., they are
not aimed at servicing for real commodity circulation, but are performed
to gain profits from the scheme: "bought cheaper and sold better". They
all are based on the transactors' predictions about rate changes made.
At the same time, and it is very important, the predictions made by the
participants of each transaction are polar. So the volume of speculative
transactions characterizes the measure of discrepancies in the market
participants' predictions, i.e., in the reality, the measure of
unpredictability of financial time-series.
This most important property of market time-series underlies the efficient
market hypothesis put forth by Louis Bachelier in his thesis in
1900. According to this doctrine, an investor can only hope for the
average market profitableness estimated using such indexes as Dow Jones
or S&P500 (for New York Exchange). However, every speculative profit
occurs at random and is similar to a gamble. The unpredictability of
market curves is determined by the same reason for which money can be
hardly found lying on the ground in busy streets: there are too many
volunteers to pick it up.
The efficient market theory is not supported, quite naturally, by the
market participants themselves (because they are exactly in search of
this "lying" money). Most of them are sure that market time-series,
though they seem to be stochastic, are full of hidden regularities,
i.e., they are at least partly predictable. It was Ralph Elliott, the
founder of technical analysis, who tried to discover such hidden
empirical regularities in the 30's.
In the 80's, this point of view found a surprising support in the
dynamical chaos theory that had occurred shortly before. The theory is
based on the opposition of chaos state and stochasticity (randomness).
Chaotic series only appear random, but, as a determined dynamical
process, they leave quite a room for
a short-term forecast. The area of feasible forecasts
is limited in time by the forecasting horizon, but that may be
sufficient to gain real profits due to forecasting (Chorafas, 1994).
Then those having better mathematical methods of extracting regularities
from noisy chaotic series may hope for a better profit rate - at the
expense of their worse equipped fellows.
In this article, we will give specific facts confirming the partial
predictability of financial time-series and even numerically evaluate
Author: MetaQuotes Software Corp.
This is a work of art. After all the writing, they concluded: "We demonstrated that (at least some of) market time-series were partly predictable."
At least some? Take any time frame of one hour, tag the start price and end price, enter a position in the direction of the price and it will be right for at least a few ticks at least 50% of the time. Any bot is only as good as it's programming logic, just as any humans logic is. I can also curve fit a bot to do anything, and a self learning bot can only "learn" from it's experience of "life". I totally agree a bot can do wonderful things and I myself use them. What we should remember though is, bot or human, neither can predict the future (certainly not a bot at least), but only make an educated guess as to what might happen based on past experience. Instead of saying "predicting the future", we should say "recognizing patterns that have behaved in a certain way in the past". I know it's easier to say it the first way, but also misleading. The idea of using a bot to learn the best patterns is good, I'm just suggesting to say it as it is, without all the juggling of words.
Hi Metaquotes, would it be possible if you could ask Mr. Sergey Shumskiy, these comments and questions:
Thank you very much. The article was very insightful and helpful.
1. Could you further explain the Box-Counting Method? This is very confusing.
2. What do you mean by "information dimension of increment" above Fig 4?
3. How did you create the fig 4? What program and tools did you use? I am having a hard time understanding your logic on the empirical reasoning for financial times series' predictability.
4. How did you create fig 5? And is this a method of choosing the optimial immersion depth?
interpret the cross-entropy eqn? (what are H and delta?) What is cross-entropy?
5. You mentioned in NN learning, there's training, validating, and testing. How are validating and testing different?
What is the orthogonalization technique? What
are the relationship between dimensional lag space, main components, and
attributes (Method of Hints)
Thank you and I hope Sergey can somehow get this feedback because I'd love to hear his responses.