Review of time series forecasting models: pen test

In the framework of my thesis, “ Prediction Model for a Sample of Maximum Similarity, ” I needed to review forecast models . In addition to the review, I made a classification option, which I did not really succeed then. The classification has already been slightly corrected, now I want to understand the existing models for forecasting time series. Such models are called stochastic models.

According to one Tikhonov in his " Forecasting in market conditions " today (2006) there are about 100 methods and models of forecasting. This assessment sounds crazy, I fully dismantled it ! Let's now look at what kind of time series forecasting models exist today.

1. Regression prediction models
2. Autoregressive Prediction Models (ARIMAX, GARCH, ARDLM)
3. Exponential Smoothing Models (ES)
4. Maximum Sampling Model (MMSP)
5. Model on neural networks (ANN)
6. Model on Markov chains
7. Model on classification regression trees (CART)
8. Model based on genetic algorithm (GA)
9. Model on support vectors (SVM)
10. Model based on transfer functions (TF)
11. Fuzzy Logic Model (FL)
12. What else?...

Regression prediction models

Regression prediction models are one of the oldest, but one cannot say that it is very popular now. Regression models are:
')

• Simple linear regression
• Multiple regression (multilple regression)
• Nonlinear regression

The best regression book is the archigenic book — Draper N., Smith H. Applied regression analysis . It can be downloaded online in djvu. It is better to read in the English version, written in the highest degree for people.

Autoregressive Prediction Models

This is the broadest and one of the two most widely applicable classes of models! There are many books on these models, many examples of use.

• ARIMAX (autoregression integrated moving average average extended), this is written very much. The basics of the book are Box, George and Jenkins, Gwilym (1970) Time series analysis: Forecasting and control. Better to read in English!
• GARCH (generalized autoregressive conditional heteroskedasticity), there are many modifications of FIGARCH, NGARCH, IGARCH, EGARCH, GARCH-M.
• ARDLM (autoregression distributed lag model), about this only in econometrics textbooks.

Question to the audience: advise a good and understandable (!) Book / article on GARCH and MLE.

Exponential Smoothing Models

• Exponential smoothing (exponential smoothing)
• Holt model or double exponential smoothing
• Holt Winters model or triple exponential smoothing

For all three models, the best article I've read is Prajakta SK Time series Forecasting using Holt-Winters Exponential Smoothing .

Model by sampling maximum similarity

This is my model (model on the most similar pattern), on a number of tasks it shows high efficiency. To the ranks of FOREX and exchanges do not apply, checked, it does not work. Its description can be found in the thesis on the link above, in addition, you can download an example implementation in MATLAB .

Model on neural networks

The second of the two most popular time series forecasting models. The best book with examples, for my taste, Khaikin S. Neural networks: a full course . An example book in MATLAB can be downloaded here.

Model on Markov's chains

The model on Markov's chains appears in a variety of reviews, but I did not manage to find either a good book or a good article about its specific application for predicting time series. This model itself was analyzed in the course of the theory of reliability (Gnedenko’s textbook), the principle of its calculation is well understood, moreover, I read that it is often used to model financial time series.

Question to the audience: advise a good and understandable (!) Book / article on the use of Markov chains to predict time series.

Model on classification regression trees

Here are some materials, but they are there. In particular, a good article on the application of this model for prediction of the Vulnerability to Famine and Chronic Food Insecurity .

Model based on genetic algorithm

This is a strange beast, I call such decisions “Jesuit”, because it seems that they were born only to substantiate scientific novelty, but their effectiveness is low. For example, a genetic algorithm is used to solve optimization problems (search for extremum), but some have dragged it into time series prediction. I did not manage to find intelligible material on this topic.

Question to the audience: advise a good and understandable (!) Book / article on the application of the genetic algorithm to predict time series.

Model on support vectors

Model based on transfer functions

Model on fuzzy logic

All these models belong, for my taste, to the Jesuit class. For example, reference vectors (SVM) are used mainly for classification problems. Fuzzy logic, where it does not apply, however, I did not succeed in finding its clearly described use for predicting time series. Although the reviews experts almost always indicate it.

The question to the audience is the same!

Total

We will get a dozen models, with all the modifications - two dozen. I wish that in the comments you not only expressed an opinion, but if possible, made useful links to understandable materials. Better in english!

Ps. All lovers of FOREX and all sorts of exchanges a big request is not a ditch to me in a personal! I'm terribly tired of you!