Nonparametric Pattern Recognition Research Articles

Price discovery redux—Analyzing energy spot and futures prices using a dynamic programming approach

We employ dynamic time warping (DTW), a non-parametric pattern recognition technique based on a dynamic programming algorithm, to analyze whether futures markets for crude oil and natural gas have facilitated price discovery over the last decade. Should futures prices absorb and reflect information before spot prices, they will move first and lead spot prices, suggesting that they dominate spot prices and play an important role in price discovery. The results show that natural gas futures prices led spot prices more frequently between 2019 and 2023 than during the five years preceding this turbulent period. In the case of crude oil, however, futures prices lagged spot prices more often than leading them. The evidence suggests that futures prices have not consistently fulfilled their price discovery role in the two energy markets. The results also indicate that short-term futures contracts play a more dominant role in price discovery than long-term contracts. Additionally, we demonstrate the advantages of DTW: it lends itself well to analyzing small samples with different orders of integration; it can discover linear and nonlinear relationships between time series; notably, it can detect period-to-period changes in the duration and direction of lead-lag associations between two series and present the results intelligibly.

An investigation of the price discovery role of futures markets: A dynamic time warping analysis of the United States corn markets

AbstractFutures markets are critical to price discovery and often dominate spot markets. We analyze the linkages between daily corn futures and spot prices in the United States using dynamic time warping. This nonparametric pattern recognition technique has several advantages over traditional time series methods. First, it can detect multiple changes in the lead‐lag associations between the two prices within short intervals; the duration with which one series leads or lags another is not assumed to be fixed. Second, the method can be applied to time series without regard to their stationarity properties. This greatly expands the scope of this method to accommodate a wide range of time series. Third, it lends itself well to studying small samples, which econometricians encounter routinely. Fourth, the results are presented intelligibly using intuitive visualizations. Our results show that futures markets are critical to price discovery; nevertheless, spot markets dominate futures markets intermittently. We discuss the results in detail, setting them in the proper context. [EconLit Citations: C14, C32, Q02, Q11].

On the linkages between energy and agricultural commodity prices: A dynamic time warping analysis

We use dynamic time warping, a non-parametric pattern recognition method, to study interlinkages between major energy and agricultural commodity prices. Cluster analysis is conducted to group commodity prices based on their behavioral likeness by maximizing the differences between groups while minimizing the differences within groups. Two clusters emerge: one comprises the prices of crude oil and six major agricultural commodities, whereas the other contains coal and natural gas prices. Regarding lead-lag associations, oil prices generally lag crop prices; however, there are periods during which the former lead the latter. Furthermore, the duration with which oil prices lead or lag crop prices changes frequently.

Asset-return momentum prediction through pattern recognition

Asset-price momentum effects are important to study in finance. A novel nonparametric pattern-recognition algorithm is proposed to identify intraday momentum based on price patterns. Specifically, a classification-regression model for predicting future price-movement patterns is designed by combining Bayesian analysis and a nonparametric density-estimation method (with different similarity measures). The model is applied to asset prices in both original and compressed forms through nonlinear dimension-reduction techniques. Regardless of the data form, this model’s predictive power is statistically and empirically significant. The model most effectively predicts asset prices with high levels of long-range dependence. Empirical analysis demonstrates that the intraday-momentum effect can occur over longer time horizons than the half-hour documented in the literature. The intraday-trading strategy based on the similarity-prediction model achieves an increased expected return and a Sharpe ratio of 11% and 0.7, respectively, compared to alternative benchmark models.

A two-state Markov chain model for slug flow in horizontal ducts

Intermittent flows are common flow patterns in gas–liquid horizontal flow and attract attention and great research effort due to its importance for industrial and engineering applications. The slug flow is typically modelled based on a unit cell varying from an elongated air bubble with a liquid film in segregated flow pattern and an aerated liquid plug, the slug region, with remarkable stochastic characteristics of its alternating regions. In this paper, a two-state Markov chain model is proposed to represent the stochastic dynamics of developed slug flow in horizontal pipes. Each state represents either the liquid slug or the elongated bubble regions and the transition probabilities dictate the change of the given discrete time measurement to stay at a given state or change. This simple but insightful description of the phenomenon allows an analytical treatment of the statistics of Markov chain stochastic process. Measurement stations with two double wire resistive sensors are used to obtain the void fraction time series and a corresponding two-state representation. Each state is classified by a threshold defined by an unsupervised and non-parametric pattern recognition approach. The state transition matrix is then estimated for each corresponding experimental point. Thus random samples can be synthetically regenerated with the same statistical features of the corresponding measurements. It is shown that the Markov chain model can successfully represent second-order statistics of the measurement, such as the autocorrelation and power spectral density, given an appropriate choice of the chain order. Moreover, a cross correlation-based time delay estimation is proposed for the two-sensor measurement station and additional information is obtained from the slug flow. Subsequently, statistics of some slug flow features are estimated using the proposed approach and their interpretation as random variables derived from the void fraction stochastic process is discussed.

Application of a Nonparametric Pattern Recognition Algorithm to the Problem of Testing the Hypothesis of the Independence of Variables of Multi-Valued Functions

Application of a Nonparametric Pattern Recognition Algorithm to the Problem of Testing the Hypothesis of the Independence of Variables of Multi-Valued Functions

Information system of remote sensing data analysis

The information system is designed to solve the problems of analyzing remote sensing data and making decisions using non-parametric pattern recognition algorithms, the synthesis of which is based on kernel estimates of probability density. Initial information is compiled by training samples on the states of the investigated natural objects and their corresponding spectral data, which are presented in Microsoft Excel files. The application of the software provides the following functions: primary processing of remote sensing data to evaluate the main quantitative characteristics of the laws of distribution of random variables; testing of the hypothesis on distributions of multivariate random variables based on non-parametric pattern recognition algorithms; estimating the state of natural objects by spectral data; mapping of territories by state of natural objects based on spectral data of earth surface elements and their corresponding spatial coordinates. Results of the presented functional tasks are saved in the form of reports in graphic files and spreadsheets of Microsoft Excel. Software tools can be used in the study of various nature objects from remote sensing data.

Application of a nonparametric pattern recognition algorithm in the problem of testing the hypothesis about the independence of variables of ambiguous functions

Application of a nonparametric pattern recognition algorithm in the problem of testing the hypothesis about the independence of variables of ambiguous functions

Formation of Sets of Independent Components of a Multidimensional Random Variable Based on a Nonparametric Pattern Recognition Algorithm

Formation of Sets of Independent Components of a Multidimensional Random Variable Based on a Nonparametric Pattern Recognition Algorithm

A method of sequentially generating a set of components of a multidimensional random variable using a nonparametric pattern recognition algorithm

We study in which way a priori information on the independence of random variables affects the approximation accuracy of a nonparametric estimate of the Rosenblatt–Parzen probability density. A new technique for generating sets of independent components of a multidimensional random variable is proposed. The methodology is based on testing the hypotheses of the independence of combinations of the multidimensional random variable components using a two-alternative nonparametric kernel algorithm for pattern recognition corresponding to the maximum likelihood criterion. Classes correspond to the domains of definition of the probability densities of sets of independent and dependent components of the multidimensional random variable. Nonparametric statistics of the kernel type are used to estimate the probability densities. The choice of the bandwidths of the kernel estimates of the probability densities is made from the condition of the minimum root-mean-square criterion. The sequential procedure for generating a set of independent components begins with the analysis of paired combinations of components of a multidimensional random variable. For each pair of components, the probability of an error in recognizing classes corresponding to the assumptions of independence and dependence of the considered components is estimated. A pair of components with the maximum difference between these errors is determined. If the errors obtained do not differ significantly, then there are no independent components in the considered multivariate random variable. If there is a significant difference in the probability estimates of class recognition errors, a pair of independent components is established. These components are included in a three-component set of a multidimensional random variable. The analysis of their combinations is carried out in the same way, following the above-described procedure. The process of generating the set of independent components is stopped when no reliable difference occurs any more between the probabilities of errors in recognizing situations belonging to the accepted classes. In this case, the previous set of independent components is the desired result. In contrast to the traditional methodology based on the Pearson criterion, the proposed approach allows us to bypass a problem of the decomposition of the range of values of random variables into multidimensional intervals. The method of generating a set of independent components of a multidimensional random variable is illustrated by the results of the analysis of spectral features of remote sensing data of forest tracts using space imagery from the Landsat-8 satellite.

Nonparametric pattern recognition algorithm for testing a hypothesis of the independence of random variables

A new method for testing a hypothesis of the independence of multidimensional random variables is proposed. The technique under consideration is based on the use of a nonparametric pattern recognition algorithm that meets a maximum likelihood criterion. In contrast to the traditional formulation of the pattern recognition problem, there is no a priori training sample. The initial information is represented by statistical data, which are made up of the values of a multivariate random variable. The distribution laws of random variables in the classes are estimated according to the initial statistical data for the conditions of their dependence and independence. When selecting optimal bandwidths for nonparametric kernel-type probability density estimates, the minimum standard deviation is used as a criterion. Estimates of the probability of pattern recognition error in the classes are calculated. Based on the minimum value of the estimates of the probabilities of pattern recognition errors, a decision is made on the independence or dependence of the random variables. The technique developed is used in the spectral analysis of remote sensing data.

Formation of sets of independent components of a multidimensional random variable based on a nonparametric pattern recognition algorithm

The possibility of circumventing the problem of decomposition of the range of values of random variables when testing various hypotheses is considered. A brief review of the literature on this problem is given. A method for forming sets of independent components of a multidimensional random variable is proposed, based on hypotheses testing about the independence of paired combinations of components of a multidimensional random variable. The method uses a two-dimensional non-parametric algorithm for pattern recognition of the kernel type, corresponding to the criterion of maximum likelihood. In contrast to the traditional method based on the application of the Pearson criterion, the proposed approach avoids the problem of decomposing the range of values of random variables into multidimensional intervals. The results of computational experiments performed according to the method of forming sets of independent random variables are presented. Using the information obtained, an information graph is constructed, the vertices of which correspond to the components of a multidimensional random variable, and the edges determine their independence. Then the vertices of the complete subgraphs correspond to groups of independent components of a random variable. The obtained results form the basis for the synthesis of a multi-level nonparametric large volume data processing system, each level of which corresponds to a specific set of independent random variables.

Nonparametric “anti-Bayesian” quantile-based pattern classification

Parametric and nonparametric pattern recognition have been studied for almost a century based on a Bayesian paradigm, which is, in turn, founded on the principles of Bayes theorem. It is well known that the accuracy of the Bayes classifier cannot be exceeded. Typically, this reduces to comparing the testing sample to mean or median of the respective distributions. Recently, Oommen and his co-authors have presented a pioneering and non-intuitive paradigm, namely that of achieving the classification by comparing the testing sample with another descriptor, which could also be quite distant from the mean. This paradigm has been termed as being “anti-Bayesian,” and it essentially uses the quantiles of the distributions to achieve the pattern recognition. Such classifiers attain the optimal Bayesian accuracy for symmetric distributions even though they operate with a non-intuitive philosophy. While this paradigm has been applied in a number of domains (briefly explained in the body of this paper), its application for nonparametric domains has been limited. This paper explains, in detail, how such quantile-based classification can be extended to the nonparametric world, using both traditional and kernel-based strategies. The paper analyzes the methodology of such nonparametric schemes and their robustness. From a fundamental perspective, the paper utilizes the so-called large sample theory to derive strong asymptotic results that pertain to the equivalence between the parametric and nonparametric schemes for large samples. Apart from the new theoretical results, the paper also presents experimental results demonstrating their power. These results pertain to artificial data sets and also involve a real-life breast cancer data set obtained from the University Hospital Centre of Coimbra. The experimental results clearly confirm the power of the proposed “anti-Bayesian” procedure, especially when approached from a nonparametric perspective.

Methods for Rapid Selection of Kernel Function Blur Coefficients in a Nonparametric Pattern Recognition Algorithm

A fast algorithm is proposed for choosing the coefficients of blur coefficients for kernel functions in a nonparametric estimate of the separating surface equation for a two-alternative pattern recognition problem. The algorithm is based on the results of a study of the asymptotic properties of nonparametric estimates of the decision function in the recognition problem for patterns and the probability densities of the distribution of random variables in classes. We compare the proposed algorithm with the traditional approach based on minimizing the estimated probability of a classification error.

Method of fast algorithm bandwidth selection for kernel function in nonparametric pattern recognition algorithm

Method of fast algorithm bandwidth selection for kernel function in nonparametric pattern recognition algorithm

Research of efficiency of the statistical non-parametric pattern recognition models for forest land classification

The principles of the creation of pattern recognition models, based on using multispectral imagery of forest land, have been analyzed. The statistical non-parametric model has been suggested as a basic pattern recognition model and the probability density function - as a recognition feature. Efficiency of the different quality criteria has been discussed. The main directions for improving the pattern recognition models are regarded.

Nonparametric Algorithms for Estimating the States of Natural Objects

Modifications of a nonparametric pattern recognition algorithm corresponding to the maximum likelihood criterion with additional decision functions are considered. The synthesis of the proposed algorithms is based on the analysis of the ratios of the estimates of the probability density distributions of random variables in classes and their functionals with input thresholds. The choice of the thresholds is determined by specific features of the classification problem. The results obtained are applied for assessing the states of forest tracts on the basis of remote sensing data.

Dynamic analysis environment for nuclear forensic analyses

A Dynamic Analysis Environment (DAE) software package is introduced to facilitate group inclusion/exclusion method testing, evaluation and comparison for pre-detonation nuclear forensics applications. Employing DAE, the multivariate signatures of a questioned material can be compared to the signatures for different, known groups, enabling the linking of the questioned material to its potential process, location, or fabrication facility. Advantages of using DAE for group inclusion/exclusion include built-in query tools for retrieving data of interest from a database, the recording and documentation of all analysis steps, a clear visualization of the analysis steps intelligible to a non-expert, and the ability to integrate analysis tools developed in different programming languages. Two group inclusion/exclusion methods are implemented in DAE: principal component analysis, a parametric feature extraction method, and k nearest neighbors, a nonparametric pattern recognition method. Spent Fuel Isotopic Composition (SFCOMPO), an open source international database of isotopic compositions for spent nuclear fuels (SNF) from 14 reactors, is used to construct PCA and KNN models for known reactor groups, and 20 simulated SNF samples are utilized in evaluating the performance of these group inclusion/exclusion models. For all 20 simulated samples, PCA in conjunction with the Q statistic correctly excludes a large percentage of reactor groups and correctly includes the true reactor of origination. Employing KNN, 14 of the 20 simulated samples are classified to their true reactor of origination.

Employment of Kernel Methods on Wind Turbine Power Performance Assessment

A power performance assessment technique is developed for the detection of power production discrepancies in wind turbines. The method employs a widely used nonparametric pattern recognition technique, the kernel methods. The evaluation is based on the trending of an extracted feature from the kernel matrix, called similarity index, which is introduced by the authors for the first time. The operation of the turbine and consequently the computation of the similarity indexes is classified into five power bins offering better resolution and thus more consistent root cause analysis. The accurate and proper detection of power production changes is demonstrated in cases of icing, power derating, operation under noise reduction mode, and incorrect controller input signal. Finally, overviews are illustrated for parks subjected to icing and operating under limited rotational speed. The comparison between multiple adjacent turbines contributes further to the correct evaluation of the park overall performance.

Non-parametric rain/no rain screening method for satellite-borne passive microwave radiometers at 19–85 GHz channels with the Random Forests algorithm

This paper presents a novel non-parametric pattern recognition method to screen rain/no rain status for satellite-borne passive microwave radiometers in the 19–85 GHz channels. The method is based on randomized decision trees with bootstrap aggregation (Random Forests (RF) algorithm). It relies on pragmatic associations between the input features using Tropical Rainfall Measuring Mission (TRMM) Microwave Imager (TMI) calibrated brightness temperatures and precipitation radar (PR) rain/no rain information as targets. Both these instruments are carried on board the TRMM satellite. In order to develop the method, first, the 10 most significant input features are selected by using feature importance criteria through out-of-bag (OOB) statistics from a total of 17 input features. The input features include the brightness temperatures, as well as some computed signatures – polarization differences (PD), polarization-corrected temperatures (PCT), and scattering indices (SI) at in the 19–85 GHz channels. The feature selection is carried out for different types of surface terrain (ocean, land, and coast), and the selected features are then used for final RF algorithm development. During the dichotomous statistical assessment of the method against the PR rain/no rain status as ‘truth’, the presented method produced reasonable threat scores of 0.50, 0.43, and 0.39, respectively, over ocean, land, and coast surface terrains. Furthermore, the results are compared with the dichotomous scores derived by the Goddard profiling algorithm (GPROF) and, remarkably, the RF-based method corroborated better statistical scores than that of the GPROF. The presented method does not rely on any a priori information and is applicable to other passive microwave radiometers at similar frequencies.