Trading Strategies and the Cat Tournament Protocol
|
|
- Grant Reeves
- 3 years ago
- Views:
Transcription
1 M A C H I N E L E A R N I N G P R O J E C T F I N A L R E P O R T F A L L 2 7 C S CLASSIFICATION OF TRADING STRATEGIES IN ADAPTIVE MARKETS MARK GRUMAN MANJUNATH NARAYANA
2 Abstract In the CAT Tournament, agents facilitate transactions between buyers and sellers with the intention of maximizing profit from commission and other fees. The agent must find a well-balanced strategy that allows it to sign buyers and sellers to trade in its market while also maintaining the buyers and sellers that are currently subscribed to it. One approach is to classify the traders that interact with the agent with respect to the trading strategies they utilize. Although the set of trading strategies is very small (traders are assigned one of four previously-defined strategies) the data available to the agent does not provide an explicit correlation of strategies to traders during the competition. As a result, agents must gather data in which the relationship between agents and strategies is known and then construct a probabilistic model based on the acquired data. Different strategies yielded varying frequency and quantity of data, making its raw form unusable for the k-means clustering technique. Only after discarding most of the data (thus deserting its integrity) was it possible to pass it to the k-means algorithm, which performed very poorly. A support vector machine (SVM) was also utilized and yielded varying results. While the linear and sigmoid kernels performed nominally better than classification based on complete randomness (28.2% and 32.8% respectively), the SVM employing a radial kernel was able to predict strategies with nearly 6% accuracy. Classification using a Hidden Markov Model was even more successful, predicting strategies correctly 62.28% of the time. It was observed that a moderate number of states and mixtures yielded the best results when classifying using the HMM. Additionally, two of the four strategies were observed to be more easily predictable using both the SVM and HMM classification techniques, suggesting that better classification could be achieved if the raw data from the two strategies could be separated into two disjoint sets before performing classification. 1. Introduction The field of Catallactics, or the study of exchanges, has received significant attention in the Artificial Intelligence community over the past few years, in large part, due to the progress that has been made in Machine Learning. In particular, significant attention has been given to designing efficient markets in which agents of numerous roles and preferences interact and exchange goods while utilizing various strategies and trading tactics. The CAT Tournament [1], an offshoot of the original Trading Agents Competition, is such a contest in which agent brokers (hereon referred to as specialists ) attempt to lure buyers and sellers (hereon referred to as traders ) to their respective markets in hopes of maximizing profit. A variety of trading techniques employed by the traders, who also wish to maximize individual profit, serve as the main motivation for developing adaptive markets that actively respond to the ever-changing preferences of the buyers and sellers. In addition to enticing traders to utilize the specialist s unique market, the specialist must also make necessary adjustments to its strategy to retain the traders that have decided to collaborate with it. Inevitably, trade-offs must be made regarding whether to maintain a similar strategy while possibly limiting the amount of profit that could be made and altering strategy to maximize profit at the risk of losing existing traders. Therefore, in order to make optimal decisions, the specialist must make inferences about the expected behavior of traders given any action it may take. To do this, the specialist must analyze data it has collected from previous trading sessions to model the strategies employed by the traders. To simplify this task, the CAT protocol specifies that traders are most likely (and possibly required) to use one of four previously-defined trading strategies, turning the learning problem (at least in part) into a clustering problem. Naturally, the relevance of said clustering also comes into question. In other words, one must decide whether an ability to classify all traders based on the previously-defined strategies is useful in maximizing the overall profit of the specialist agent. Intuitively, successful classification should allow the specialist to make better inferences regarding the overall effect of potential change in trading strategy, although the degree to which the inferences are improved can only be determined through research. 2
3 2. Strategies The traders in the CAT tournament are designed to choose from one of four strategies, which are explained in detail in [2]-[5]. An agent using the Double Auction strategy [2] (henceforth called ) keeps track of the number of bids accepted and rejected by the market at a particular price. Subsequent values for the bids are chosen depending on the probability of acceptance of a bid, given the past history. Utilizing the Extensive Form Game strategy [3] (hereon referred to as ), a trader alters its future bid values based on the profits that were observed for the previous bids. The Zero Information-Constrained (ZI-C) strategy [4] involves generating random bids constrained between a maximum and minimum value. A buyer using the ZI- C strategy will never bid more than what it believes a good is worth. Likewise, a seller will never sell a good for less than the amount it cost the seller to obtain the good. Finally, if a trader uses the Zero Information, Plus () strategy [5], it utilizes the same trading techniques as an agent that employs ZI-C, but it also updates the constraints based on feedback from the market. Thus, each process (except when using the ZI-C strategy) receives feedback from the market in various forms and updates itself to generate new bids. 3. Data Each trader makes a bid to the market and continues to update it until another trader in the same market accepts the bid price and a transaction takes place. We call each string of updated bids from the same trader a bid sequence. Given the competitive nature of the market with several traders attempting to make the transaction, a buyer s bid could be accepted by a seller at any point during the bid process. This means that the number of bids a trader had to make before successfully concluding a transaction is not constant. As a result, the number of bids in each bid sequence can vary significantly. There is no upper bound on the length of the sequence. Figure 1, illustrating bid sequences for a selling trader using the strategy, shows that the number of sequences in the bid could range from 1 to any large number depending on the state of the market and the strategies of other traders in that market. The problem is further complicated when we consider multiple traders using different strategies. Figure 2 is an illustration of how the bid sequences might look in the actual competition. We can see that the data clearly poses many challenges, the most significant being the unpredictable lengths of bid sequences. Seller 1-84., 9.4, 12.9 Seller Seller Seller 1-59., 85.2, 82.7, 81.3, 73.4, Figure 1. Illustration of bid sequences from one seller using the strategy. Seller 1-84., 9.4, 12.9 Seller Seller 3-3.9, 14.5, 75.6, 9.3 Seller 1-1.2, 98.7, 89.6, 77.6, 19.4 Seller 4-59., 85.2, 82.7, 81.3, 73.4 Seller , 56.7, 78.9 Seller 3-2.8, 132,6 Figure 2. Illustration of bid sequences from a sample market with multiple traders employing different bidding strategies. The observed data is further distorted by the processes of generating new bids. As explained in section 2, the underlying parameters of each bidding strategy are updated using market feedback. Each trader gets different feedback from the market depending on the values of its previous bid and underlying parameters. Thus, the parameters of traders using even the same strategy are updated differently. Consequently, we see differing bid prices from traders using the same strategy. 4. Data Collection The traders bidding data was collected through numerous simulations of a typical CAT competition. Although there were no problems collecting a sufficient amount of data (one could always run more simulations), the raw data collected could not immediately be used for classification purposes for a number of reasons. The most significant obstacle of data classification was dealing with anonymized data. By anonymized we mean that the true origin of each bid was masked. This occurred per the 2
4 specification of the CAT Competition Protocol, in which bids shouted by traders first reached the server, which replaced the identity of the bid source with a unique bid identifier used only for that particular sequence of bids. Once a transaction completed, the bid identifier was discarded and a new bid identifier was assigned to the next sequence of bids from the same trader. As a result, all bids were masked by the time they reached the specialist. Thus, the specialist could determine neither the true identity of each bid, nor which bid sequences originated from the same trader. The random order in which the bids arrived also failed to assist the specialist in determining the true origin of each bid. Fortunately, the CAT source code (freely available to all CAT Competition participants) gives users access to all functional modules that make up the competition. With these additional resources, and a number of code modifications, the authors were able to obtain the required data in its unmasked form, allowing them to identify the true source of each bid that the specialist received. This data was collected under the assumption that it would be used for training only, since the identity of each bid would not be available to the specialist during an actual competition. Bid sequences were collected for 4 traders (1 traders of each strategy), with their identities unmasked. The authors decided that focusing on the selling trader sequences alone would suffice to evaluate the efficacy of the classification strategy. In all, 276 bid sequences were generated by the system. Two-thirds (1384) of these samples were randomly chosen to be the training set and one-third of the data (692) was used for testing. 5. Classification A number of classification techniques were employed to determine which technique yielded the best classification results. We discuss in detail the techniques that showed the most promising results and touch upon other techniques that were evaluated in limited capacity. I. K-means classification: The first classification method utilized the k-means classifier in MATLAB. In order to use this classification technique, the collected data needed to be adjusted to ensure that each data entry contained the same number of data points. As a result, a decision was made to include only the first bid of each recorded bid sequence. This decision was made on the observation that some (in fact a large majority of) bid sequences were made up of only a single bid, so only the first one could be used in order to guarantee that 1) each bid sequence was represented in the data, and 2) all four strategies were equally represented. Figure 3 illustrates the initial plot of the original raw data. The k-means clustering algorithm created 4 clusters that completely disagreed with the original raw data plot, as seen in Figure 4. This result was very intuitive, given the seemingly random plot of the original raw data which lacked any identifiable clusters, and suggested that k-means clustering should not be used to classify testing data. Figure 3. Each bid sequence is represented by the first bid. In all, 135 bid sequences of each strategy are represented. Although discarding all bids except the first from each bid sequence guaranteed equal representation of all four trading strategies, it compromised the integrity of the data by removing a significant number of data points pertaining to certain bid sequences. The bid sequences that were especially long (some 3
5 made up of more than 1 bids) were affected most significantly by this approach. One alternative that was not tested would have involved treating each bid as its own sequence. Although this approach would avoid discarding a large number of data points, it is not intuitive whether the integrity of the data would be maintained, since the relationship of bids and the bid sequence they belonged to would be lost in the transformation. type of kernel that was used for training. Training on the sigmoid kernel (with default gamma and coefficient values) yielded the worst results, predicting only 28.2% of the testing set correctly (only 3% better than completely random prediction). Training under the linear kernel also yielded rather poor results, predicting only 32.8% of the training data correctly. The polynomial kernel also yielded rather poor results, predicting 38.4% of the testing data correctly using the default polynomial kernel. The sigmoid kernel, however, produced much better results. It was able to predict 53.8% of the testing data correctly under default parameters and 59.7% of the data correctly when gamma was set to.8. An observation was also made that the and ZI-C strategies were predicted with a high degree of accuracy, while data from the and strategies was more difficult to classify. Figure 4. Erroneous K-means clustering of the observed data. II. SVM classification: Another attempt to classify the observed data was made using a Support Vector Machine (SVM). In order to perform classification using the SVM, the observed raw data was converted to the appropriate data format. Each sequence of bids was represented by a unique vector, and each bid in the sequence became a feature in the corresponding vector. Feature numbers were assigned incrementally, so a bid sequence made up of n bids was represented by a vector of features 1 through n. Each vector was assigned a classification as follows: a bid sequence coming from a trader using the strategy was classified as class 1, as class 2, as class 3, and as class 4. An example of the data can be seen in Figure 5. SVM training and classification was performed using LibSVM v2.85. Prediction results using SVM classification varied greatly, depending largely on the Figure 5. Example of observed data in SVMfriendly format. The confusion matrix represents the quality of classification by the system. Each row represents the true class of a data instance and each column represents the predicted class of a data instance. Thus, the confusion matrix indicates the errors within each class (how many instances of each class were successfully classified and how many were misclassified). High diagonal values in the matrix indicate a successful classification system. Figure 6 presents the confusion matrices for various runs of the SVM classification with different kernels. 4
6 SVM Linear kernel Accuracy = 32.8% SVM Radial kernel Default gamma= Accuracy = 53.8% Figure 6. Confusion matrices for different SVM kernels III. HMM classification: From the onset, the presence of hidden parameters in the data (bid identities, strategies, and the internal parameters within each strategy), and time-dependent nature of bid sequences suggested that a Hidden Markov Model should be utilized in classifying the acquired dataset. Q SVM Polynomial kernel Default gamma=1, coeff=, degree= Accuracy = 38.4% SVM Sigmoid kernel Accuracy = 28.2% Hidden State Default gamma=1, coeff= the system. The hidden variables are the state (Q) and the mixture parameters (M). Since there are 4 strategies, 4 separate HMMs are trained (one for each class of data). Each HMM thus represents one of 4 strategies. For the testing set, the likelihood of each data sample for each HMM is calculated. The data sample is assigned to the strategy that is represented by the HMM with the maximum likelihood value. The HMM-based classification showed a slight improvement over the SVM-based method that was used earlier, supporting the authors initial observation of the possible effectiveness of the Hidden Markov Model in representing the variables involved in the bidding process. Several HMM runs were executed with different values for the parameters (number of hidden states and mixture components). The HMM models took between 1 and 68 minutes to train depending on the number of states and mixture components. In contrast, SVM training ranged between only a few seconds and 2-3 minutes. Nevertheless, the accuracy for the HMM method was always greater than 52%, and the authors were able to achieve about 62% accuracy by tuning the parameters (number of hidden states = 1, number of mixture components= 1). Figure 8 is a plot of the accuracies obtained in different HMM run, and Figure 9 shows the confusion matrices for a few runs of the HMM method. M 64. Classification accuracy for different HMM runs Mixture parameters X Accuracy Bid Figure 7. Mixture-observation Hidden Markov Model used for the CAT dataset. The Mixture-observation HMM [6] that was used for this dataset is shown in Figure 7. A Gaussian mixture model is assumed for each observation. X represents the current bid value and is the observed variable in States, 1 Mixture 4 States, 4 States, 4 States, 2 Mixtures 1 Mixtures 1 Mixtures HMM parameters 1 States, 1 Mixtures Figure 8. Classification accuracy for different HMM parameters. 5
7 From the HMM confusion matrices presented in this report, it is clear that the and ZI-C strategies are the easiest to classify. This fact can also be observed from figure 3, where the red dots (representing the strategy) clearly follow a recognizable pattern, and the cyan dots (representing the ZI-C strategy) all fall within a reasonably small range of values. The strategy was moderately difficult to classify, while the strategy proved extremely difficult (its performance was not much better than purely random performance). Figure 9. Confusion matrices for different HMM parameters 6. Feature reduction A small experiment was also carried out to explore the efficacy of feature reduction of the observed dataset in the HMM classification framework. We could not use standard reduction methods like PCA because the dataset included instances of varying feature lengths (treating each bid as a single feature). Most samples had only 1 feature, while some had as many as 2. Experiments showed that when only 2 features were used, the HMM accuracy fell to 49%, while using only the first 1 features improved upon earlier best results slightly. Thus, the authors concluded that a moderate reduction in the number of features could result in improved performance while also maintaining the integrity of the observed data. Further experiments would be required to arrive at the optimal number of features that should be used. The confusion matrices and accuracies for 2- feature-only and 1-features-only HMM runs are presented in Figure 1. HMM First 2 features only Accuracy = 49.71% HMM First 1 features only Figure 1. Results of training an HMM sing only first 2 and first 1 bids from each bid sequence, respectively. 7. Other methods A number of other interesting methods were explored for improving the results of the HMM. We tried to fit a Conditional Random Fields (CRF) model to the data, using Kevin Murphy s Matlab code [7]. The model failed to converge in many cases, however, and resulted in accuracy close to random performance for our dataset. We concluded that a deeper understanding of the CRF process was required before we could effectively utilize the advantages that could be offered by the discriminative model of CRF. Since SVM is the most popular and successful classification method in most applications, we decided to also try the time-dependent Fourier kernel [8]. We thought that the power of SVM when combined with some time information provided by the Fourier kernel might show significant improvement in the classification accuracies. A Fourier kernel was calculated for all the instances in the training and testing data set, as required by the libsvm framework for implementing user-defined kernels in libsvm. Accuracy of between 28% and 42% was observed for different cases of the Fourier kernel, suggesting that it was not an improvement over the other kernel functions of the SVM framework. We briefly contemplated the use of the pyramid kernel [9] because it is designed to work with datasets that have varying number of features. However, due to lack of time we decided to pursue this as future work. 8. Conclusions and Future work We found SVM classification to be reasonably accurate, with the best case accuracy of 59%. However, SVM by nature does not handle time-series Accuracy = 62.57%
8 data well and is also adversely affected by data containing varying feature lengths. The HMM proved to be a better solution for this dataset (yielding a prediction accuracy of almost 63% in the best case), because it is designed for time-series data and not affected greatly by the varying lengths of features. We used a simple HMM model for this project. It might be useful to experiment with more complicated models that represent the data generation process more accurately. Given that the task was classification of the data into multiple classes, we were surprised by the prediction performance obtained using the SVM and HMM techniques. Random guesses would yield about 25% accuracy). The achieved results were better than expected, suggesting that the classification of trading strategies as a pre-step to managing the transactions between traders in the CAT virtual market is a reasonable approach for any CAT participant. The data generated for this project was from one trading session alone. Further rigorous testing for robustness across trading sessions should be carried out before applying the classification scheme in the CAT Tournament. [6] Jeff A. Bilmes. What HMMs can do. IEICE Transactions on Information and Systems, E89-D(3): , 26. [7] crf.html [8] SVM kernels for time series analysis, Stephen Ruping, 21. [9] K. Grauman and T. Darrell. The Pyramid Match kernel: discriminative classification with sets of image features. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Beijing, China, October References [1] Cliff et al. CAT Document 1. TAC Market Design: Communication Protocol Specification. July 27. [2] Dave Cliff. Minimal-intelligence agents for bargaining behaviours in market-based environments. Technical Report HP-97-91, Hewlett-Packard Research Laboratories, Bristol, England, [3] S. Gjerstad and J. Dickhaut. Price formation in double auctions. Games and Economic Behaviour, 22:1-29, [4] D. K. Gode and S. Sunder. Allocative efficiency of markets with zero-intelligence traders: Markets as a partial substitute for individual rationality. The Journal of Political Economy, 11(1): , February [5] A. E. Roth and I. Erev. Learning in extensiveform games: Experimental data and simple dynamic models in the intermediate term. Games and Economic Behavior, 8: ,
Multi-Agent based simulation of FOREX exchange market
Multi-Agent based simulation of FOREX exchange market Vivien Delage a b, Christian Brandlhuber b, Karl Tuyls a, and Gerhard Weiss a a Department of Knowledge Engineering, Maastricht University, the Netherlands
More informationEmployer Health Insurance Premium Prediction Elliott Lui
Employer Health Insurance Premium Prediction Elliott Lui 1 Introduction The US spends 15.2% of its GDP on health care, more than any other country, and the cost of health insurance is rising faster than
More informationImpact of Misalignment of Trading Agent Strategy across Multiple Markets
Impact of Misalignment of Trading Agent Strategy across Multiple Markets Jung-woo Sohn, Sooyeon Lee, and Tracy Mullen College of Information Sciences and Technology, The Pennsylvania State University,
More informationSocial Media Mining. Data Mining Essentials
Introduction Data production rate has been increased dramatically (Big Data) and we are able store much more data than before E.g., purchase data, social media data, mobile phone data Businesses and customers
More information1 Topic. 2 Scilab. 2.1 What is Scilab?
1 Topic Data Mining with Scilab. I know the name "Scilab" for a long time (http://www.scilab.org/en). For me, it is a tool for numerical analysis. It seemed not interesting in the context of the statistical
More informationMaking Sense of the Mayhem: Machine Learning and March Madness
Making Sense of the Mayhem: Machine Learning and March Madness Alex Tran and Adam Ginzberg Stanford University atran3@stanford.edu ginzberg@stanford.edu I. Introduction III. Model The goal of our research
More informationBIOINF 585 Fall 2015 Machine Learning for Systems Biology & Clinical Informatics http://www.ccmb.med.umich.edu/node/1376
Course Director: Dr. Kayvan Najarian (DCM&B, kayvan@umich.edu) Lectures: Labs: Mondays and Wednesdays 9:00 AM -10:30 AM Rm. 2065 Palmer Commons Bldg. Wednesdays 10:30 AM 11:30 AM (alternate weeks) Rm.
More informationMAXIMIZING RETURN ON DIRECT MARKETING CAMPAIGNS
MAXIMIZING RETURN ON DIRET MARKETING AMPAIGNS IN OMMERIAL BANKING S 229 Project: Final Report Oleksandra Onosova INTRODUTION Recent innovations in cloud computing and unified communications have made a
More informationForecasting Trade Direction and Size of Future Contracts Using Deep Belief Network
Forecasting Trade Direction and Size of Future Contracts Using Deep Belief Network Anthony Lai (aslai), MK Li (lilemon), Foon Wang Pong (ppong) Abstract Algorithmic trading, high frequency trading (HFT)
More informationComparing the Results of Support Vector Machines with Traditional Data Mining Algorithms
Comparing the Results of Support Vector Machines with Traditional Data Mining Algorithms Scott Pion and Lutz Hamel Abstract This paper presents the results of a series of analyses performed on direct mail
More informationBEHAVIOR BASED CREDIT CARD FRAUD DETECTION USING SUPPORT VECTOR MACHINES
BEHAVIOR BASED CREDIT CARD FRAUD DETECTION USING SUPPORT VECTOR MACHINES 123 CHAPTER 7 BEHAVIOR BASED CREDIT CARD FRAUD DETECTION USING SUPPORT VECTOR MACHINES 7.1 Introduction Even though using SVM presents
More informationComponent Ordering in Independent Component Analysis Based on Data Power
Component Ordering in Independent Component Analysis Based on Data Power Anne Hendrikse Raymond Veldhuis University of Twente University of Twente Fac. EEMCS, Signals and Systems Group Fac. EEMCS, Signals
More informationMultiple Kernel Learning on the Limit Order Book
JMLR: Workshop and Conference Proceedings 11 (2010) 167 174 Workshop on Applications of Pattern Analysis Multiple Kernel Learning on the Limit Order Book Tristan Fletcher Zakria Hussain John Shawe-Taylor
More informationAzure Machine Learning, SQL Data Mining and R
Azure Machine Learning, SQL Data Mining and R Day-by-day Agenda Prerequisites No formal prerequisites. Basic knowledge of SQL Server Data Tools, Excel and any analytical experience helps. Best of all:
More informationStock Trading by Modelling Price Trend with Dynamic Bayesian Networks
Stock Trading by Modelling Price Trend with Dynamic Bayesian Networks Jangmin O 1,JaeWonLee 2, Sung-Bae Park 1, and Byoung-Tak Zhang 1 1 School of Computer Science and Engineering, Seoul National University
More informationStatistical Models in Data Mining
Statistical Models in Data Mining Sargur N. Srihari University at Buffalo The State University of New York Department of Computer Science and Engineering Department of Biostatistics 1 Srihari Flood of
More informationThe Scientific Data Mining Process
Chapter 4 The Scientific Data Mining Process When I use a word, Humpty Dumpty said, in rather a scornful tone, it means just what I choose it to mean neither more nor less. Lewis Carroll [87, p. 214] In
More informationTowards better accuracy for Spam predictions
Towards better accuracy for Spam predictions Chengyan Zhao Department of Computer Science University of Toronto Toronto, Ontario, Canada M5S 2E4 czhao@cs.toronto.edu Abstract Spam identification is crucial
More informationCS 2750 Machine Learning. Lecture 1. Machine Learning. http://www.cs.pitt.edu/~milos/courses/cs2750/ CS 2750 Machine Learning.
Lecture Machine Learning Milos Hauskrecht milos@cs.pitt.edu 539 Sennott Square, x5 http://www.cs.pitt.edu/~milos/courses/cs75/ Administration Instructor: Milos Hauskrecht milos@cs.pitt.edu 539 Sennott
More informationStatistical Machine Learning
Statistical Machine Learning UoC Stats 37700, Winter quarter Lecture 4: classical linear and quadratic discriminants. 1 / 25 Linear separation For two classes in R d : simple idea: separate the classes
More informationRecognizing Informed Option Trading
Recognizing Informed Option Trading Alex Bain, Prabal Tiwaree, Kari Okamoto 1 Abstract While equity (stock) markets are generally efficient in discounting public information into stock prices, we believe
More informationPredict Influencers in the Social Network
Predict Influencers in the Social Network Ruishan Liu, Yang Zhao and Liuyu Zhou Email: rliu2, yzhao2, lyzhou@stanford.edu Department of Electrical Engineering, Stanford University Abstract Given two persons
More informationPATTERN RECOGNITION AND MACHINE LEARNING CHAPTER 4: LINEAR MODELS FOR CLASSIFICATION
PATTERN RECOGNITION AND MACHINE LEARNING CHAPTER 4: LINEAR MODELS FOR CLASSIFICATION Introduction In the previous chapter, we explored a class of regression models having particularly simple analytical
More informationAnalysis of Bayesian Dynamic Linear Models
Analysis of Bayesian Dynamic Linear Models Emily M. Casleton December 17, 2010 1 Introduction The main purpose of this project is to explore the Bayesian analysis of Dynamic Linear Models (DLMs). The main
More informationMatching Rules and Market Share in an Electronic Trading Platform
Matching Rules and Market Share in an Electronic Trading Platform by Yongliang Wang A thesis presented to the University of Waterloo in fulfilment of the thesis requirement for the degree of Master of
More informationBeating the NCAA Football Point Spread
Beating the NCAA Football Point Spread Brian Liu Mathematical & Computational Sciences Stanford University Patrick Lai Computer Science Department Stanford University December 10, 2010 1 Introduction Over
More informationPredict the Popularity of YouTube Videos Using Early View Data
000 001 002 003 004 005 006 007 008 009 010 011 012 013 014 015 016 017 018 019 020 021 022 023 024 025 026 027 028 029 030 031 032 033 034 035 036 037 038 039 040 041 042 043 044 045 046 047 048 049 050
More informationAnalysis of kiva.com Microlending Service! Hoda Eydgahi Julia Ma Andy Bardagjy December 9, 2010 MAS.622j
Analysis of kiva.com Microlending Service! Hoda Eydgahi Julia Ma Andy Bardagjy December 9, 2010 MAS.622j What is Kiva? An organization that allows people to lend small amounts of money via the Internet
More informationA Platform for Trading Agent Competition
A Platform for Trading Agent Competition Dongmo Zhang and Chun Gao Intelligent Systems Laboratory University of Western Sydney Australia Outline 1 Trading Agent Competition 2 Market Design Game 3 Strategic
More informationCHARACTERISTICS IN FLIGHT DATA ESTIMATION WITH LOGISTIC REGRESSION AND SUPPORT VECTOR MACHINES
CHARACTERISTICS IN FLIGHT DATA ESTIMATION WITH LOGISTIC REGRESSION AND SUPPORT VECTOR MACHINES Claus Gwiggner, Ecole Polytechnique, LIX, Palaiseau, France Gert Lanckriet, University of Berkeley, EECS,
More informationServer Load Prediction
Server Load Prediction Suthee Chaidaroon (unsuthee@stanford.edu) Joon Yeong Kim (kim64@stanford.edu) Jonghan Seo (jonghan@stanford.edu) Abstract Estimating server load average is one of the methods that
More informationPractical Data Science with Azure Machine Learning, SQL Data Mining, and R
Practical Data Science with Azure Machine Learning, SQL Data Mining, and R Overview This 4-day class is the first of the two data science courses taught by Rafal Lukawiecki. Some of the topics will be
More informationExample: Credit card default, we may be more interested in predicting the probabilty of a default than classifying individuals as default or not.
Statistical Learning: Chapter 4 Classification 4.1 Introduction Supervised learning with a categorical (Qualitative) response Notation: - Feature vector X, - qualitative response Y, taking values in C
More informationKEITH LEHNERT AND ERIC FRIEDRICH
MACHINE LEARNING CLASSIFICATION OF MALICIOUS NETWORK TRAFFIC KEITH LEHNERT AND ERIC FRIEDRICH 1. Introduction 1.1. Intrusion Detection Systems. In our society, information systems are everywhere. They
More informationAdvanced Signal Processing and Digital Noise Reduction
Advanced Signal Processing and Digital Noise Reduction Saeed V. Vaseghi Queen's University of Belfast UK WILEY HTEUBNER A Partnership between John Wiley & Sons and B. G. Teubner Publishers Chichester New
More informationMachine Learning and Pattern Recognition Logistic Regression
Machine Learning and Pattern Recognition Logistic Regression Course Lecturer:Amos J Storkey Institute for Adaptive and Neural Computation School of Informatics University of Edinburgh Crichton Street,
More informationCompression algorithm for Bayesian network modeling of binary systems
Compression algorithm for Bayesian network modeling of binary systems I. Tien & A. Der Kiureghian University of California, Berkeley ABSTRACT: A Bayesian network (BN) is a useful tool for analyzing the
More informationData Quality Mining: Employing Classifiers for Assuring consistent Datasets
Data Quality Mining: Employing Classifiers for Assuring consistent Datasets Fabian Grüning Carl von Ossietzky Universität Oldenburg, Germany, fabian.gruening@informatik.uni-oldenburg.de Abstract: Independent
More informationSupport Vector Machines with Clustering for Training with Very Large Datasets
Support Vector Machines with Clustering for Training with Very Large Datasets Theodoros Evgeniou Technology Management INSEAD Bd de Constance, Fontainebleau 77300, France theodoros.evgeniou@insead.fr Massimiliano
More informationRethinking the Social Contract: An Agent- Based Model approach
6th International Conference on Industrial Engineering and Industrial Management. XVI Congreso de Ingeniería de Organización. Vigo, July 18-20, 2012 Rethinking the Social Contract: An Agent- Based Model
More informationSupport Vector Machines for Dynamic Biometric Handwriting Classification
Support Vector Machines for Dynamic Biometric Handwriting Classification Tobias Scheidat, Marcus Leich, Mark Alexander, and Claus Vielhauer Abstract Biometric user authentication is a recent topic in the
More informationLinear Threshold Units
Linear Threshold Units w x hx (... w n x n w We assume that each feature x j and each weight w j is a real number (we will relax this later) We will study three different algorithms for learning linear
More informationIntroduction to Logistic Regression
OpenStax-CNX module: m42090 1 Introduction to Logistic Regression Dan Calderon This work is produced by OpenStax-CNX and licensed under the Creative Commons Attribution License 3.0 Abstract Gives introduction
More informationEmissions Permits Auctions: an ABM Analysis
Emissions Permits Auctions: an ABM Analysis 1 Emissions Permits Auctions: an ABM Analysis Marta Posada, Cesáreo Hernández, Adolfo López-Paredes 1 University of Valladolid. INSISOC, Spain {posada, cesareo}@eis.uva.es
More informationRelational Learning for Football-Related Predictions
Relational Learning for Football-Related Predictions Jan Van Haaren and Guy Van den Broeck jan.vanhaaren@student.kuleuven.be, guy.vandenbroeck@cs.kuleuven.be Department of Computer Science Katholieke Universiteit
More informationIntrusion Detection via Machine Learning for SCADA System Protection
Intrusion Detection via Machine Learning for SCADA System Protection S.L.P. Yasakethu Department of Computing, University of Surrey, Guildford, GU2 7XH, UK. s.l.yasakethu@surrey.ac.uk J. Jiang Department
More informationAcknowledgments. Data Mining with Regression. Data Mining Context. Overview. Colleagues
Data Mining with Regression Teaching an old dog some new tricks Acknowledgments Colleagues Dean Foster in Statistics Lyle Ungar in Computer Science Bob Stine Department of Statistics The School of the
More informationGamma Distribution Fitting
Chapter 552 Gamma Distribution Fitting Introduction This module fits the gamma probability distributions to a complete or censored set of individual or grouped data values. It outputs various statistics
More informationBIDM Project. Predicting the contract type for IT/ITES outsourcing contracts
BIDM Project Predicting the contract type for IT/ITES outsourcing contracts N a n d i n i G o v i n d a r a j a n ( 6 1 2 1 0 5 5 6 ) The authors believe that data modelling can be used to predict if an
More informationActive Learning SVM for Blogs recommendation
Active Learning SVM for Blogs recommendation Xin Guan Computer Science, George Mason University Ⅰ.Introduction In the DH Now website, they try to review a big amount of blogs and articles and find the
More informationCrowdfunding Support Tools: Predicting Success & Failure
Crowdfunding Support Tools: Predicting Success & Failure Michael D. Greenberg Bryan Pardo mdgreenb@u.northwestern.edu pardo@northwestern.edu Karthic Hariharan karthichariharan2012@u.northwes tern.edu Elizabeth
More informationProgramming Exercise 3: Multi-class Classification and Neural Networks
Programming Exercise 3: Multi-class Classification and Neural Networks Machine Learning November 4, 2011 Introduction In this exercise, you will implement one-vs-all logistic regression and neural networks
More informationBayesian Statistics: Indian Buffet Process
Bayesian Statistics: Indian Buffet Process Ilker Yildirim Department of Brain and Cognitive Sciences University of Rochester Rochester, NY 14627 August 2012 Reference: Most of the material in this note
More informationIntroduction to Support Vector Machines. Colin Campbell, Bristol University
Introduction to Support Vector Machines Colin Campbell, Bristol University 1 Outline of talk. Part 1. An Introduction to SVMs 1.1. SVMs for binary classification. 1.2. Soft margins and multi-class classification.
More informationEM Clustering Approach for Multi-Dimensional Analysis of Big Data Set
EM Clustering Approach for Multi-Dimensional Analysis of Big Data Set Amhmed A. Bhih School of Electrical and Electronic Engineering Princy Johnson School of Electrical and Electronic Engineering Martin
More informationAuxiliary Variables in Mixture Modeling: 3-Step Approaches Using Mplus
Auxiliary Variables in Mixture Modeling: 3-Step Approaches Using Mplus Tihomir Asparouhov and Bengt Muthén Mplus Web Notes: No. 15 Version 8, August 5, 2014 1 Abstract This paper discusses alternatives
More informationlarge-scale machine learning revisited Léon Bottou Microsoft Research (NYC)
large-scale machine learning revisited Léon Bottou Microsoft Research (NYC) 1 three frequent ideas in machine learning. independent and identically distributed data This experimental paradigm has driven
More informationMachine learning for algo trading
Machine learning for algo trading An introduction for nonmathematicians Dr. Aly Kassam Overview High level introduction to machine learning A machine learning bestiary What has all this got to do with
More informationAutomated Content Analysis of Discussion Transcripts
Automated Content Analysis of Discussion Transcripts Vitomir Kovanović v.kovanovic@ed.ac.uk Dragan Gašević dgasevic@acm.org School of Informatics, University of Edinburgh Edinburgh, United Kingdom v.kovanovic@ed.ac.uk
More informationSearch Taxonomy. Web Search. Search Engine Optimization. Information Retrieval
Information Retrieval INFO 4300 / CS 4300! Retrieval models Older models» Boolean retrieval» Vector Space model Probabilistic Models» BM25» Language models Web search» Learning to Rank Search Taxonomy!
More informationBeating the MLB Moneyline
Beating the MLB Moneyline Leland Chen llxchen@stanford.edu Andrew He andu@stanford.edu 1 Abstract Sports forecasting is a challenging task that has similarities to stock market prediction, requiring time-series
More informationSegmentation and Classification of Online Chats
Segmentation and Classification of Online Chats Justin Weisz Computer Science Department Carnegie Mellon University Pittsburgh, PA 15213 jweisz@cs.cmu.edu Abstract One method for analyzing textual chat
More informationCS Master Level Courses and Areas COURSE DESCRIPTIONS. CSCI 521 Real-Time Systems. CSCI 522 High Performance Computing
CS Master Level Courses and Areas The graduate courses offered may change over time, in response to new developments in computer science and the interests of faculty and students; the list of graduate
More informationUsing Data Mining for Mobile Communication Clustering and Characterization
Using Data Mining for Mobile Communication Clustering and Characterization A. Bascacov *, C. Cernazanu ** and M. Marcu ** * Lasting Software, Timisoara, Romania ** Politehnica University of Timisoara/Computer
More informationPricing Algorithm and Network Marketing Budget
Adaptation to Dynamic Resource Availability in Ad-hoc Grids through a Learning Mechanism Behnaz Pourebrahimi, Koen Bertels Computer Engineering Laboratory, Delft University of Technology, The Netherlands
More informationT cell Epitope Prediction
Institute for Immunology and Informatics T cell Epitope Prediction EpiMatrix Eric Gustafson January 6, 2011 Overview Gathering raw data Popular sources Data Management Conservation Analysis Multiple Alignments
More informationBlind Deconvolution of Barcodes via Dictionary Analysis and Wiener Filter of Barcode Subsections
Blind Deconvolution of Barcodes via Dictionary Analysis and Wiener Filter of Barcode Subsections Maximilian Hung, Bohyun B. Kim, Xiling Zhang August 17, 2013 Abstract While current systems already provide
More informationCloud Computing. Computational Tasks Have value for task completion Require resources (Cores, Memory, Bandwidth) Compete for resources
Peter Key, Cloud Computing Computational Tasks Have value for task completion Require resources (Cores, Memory, Bandwidth) Compete for resources How much is a task or resource worth Can we use to price
More informationE-commerce Transaction Anomaly Classification
E-commerce Transaction Anomaly Classification Minyong Lee minyong@stanford.edu Seunghee Ham sham12@stanford.edu Qiyi Jiang qjiang@stanford.edu I. INTRODUCTION Due to the increasing popularity of e-commerce
More informationD-optimal plans in observational studies
D-optimal plans in observational studies Constanze Pumplün Stefan Rüping Katharina Morik Claus Weihs October 11, 2005 Abstract This paper investigates the use of Design of Experiments in observational
More informationTutorial Segmentation and Classification
MARKETING ENGINEERING FOR EXCEL TUTORIAL VERSION 1.0.8 Tutorial Segmentation and Classification Marketing Engineering for Excel is a Microsoft Excel add-in. The software runs from within Microsoft Excel
More informationThe Data Mining Process
Sequence for Determining Necessary Data. Wrong: Catalog everything you have, and decide what data is important. Right: Work backward from the solution, define the problem explicitly, and map out the data
More informationAPPM4720/5720: Fast algorithms for big data. Gunnar Martinsson The University of Colorado at Boulder
APPM4720/5720: Fast algorithms for big data Gunnar Martinsson The University of Colorado at Boulder Course objectives: The purpose of this course is to teach efficient algorithms for processing very large
More informationFOREX TRADING PREDICTION USING LINEAR REGRESSION LINE, ARTIFICIAL NEURAL NETWORK AND DYNAMIC TIME WARPING ALGORITHMS
FOREX TRADING PREDICTION USING LINEAR REGRESSION LINE, ARTIFICIAL NEURAL NETWORK AND DYNAMIC TIME WARPING ALGORITHMS Leslie C.O. Tiong 1, David C.L. Ngo 2, and Yunli Lee 3 1 Sunway University, Malaysia,
More informationFacebook Friend Suggestion Eytan Daniyalzade and Tim Lipus
Facebook Friend Suggestion Eytan Daniyalzade and Tim Lipus 1. Introduction Facebook is a social networking website with an open platform that enables developers to extract and utilize user information
More informationHandling attrition and non-response in longitudinal data
Longitudinal and Life Course Studies 2009 Volume 1 Issue 1 Pp 63-72 Handling attrition and non-response in longitudinal data Harvey Goldstein University of Bristol Correspondence. Professor H. Goldstein
More informationIntroduction to Online Learning Theory
Introduction to Online Learning Theory Wojciech Kot lowski Institute of Computing Science, Poznań University of Technology IDSS, 04.06.2013 1 / 53 Outline 1 Example: Online (Stochastic) Gradient Descent
More informationIntroduction to Machine Learning and Data Mining. Prof. Dr. Igor Trajkovski trajkovski@nyus.edu.mk
Introduction to Machine Learning and Data Mining Prof. Dr. Igor Trakovski trakovski@nyus.edu.mk Neural Networks 2 Neural Networks Analogy to biological neural systems, the most robust learning systems
More informationSupport Vector Machines Explained
March 1, 2009 Support Vector Machines Explained Tristan Fletcher www.cs.ucl.ac.uk/staff/t.fletcher/ Introduction This document has been written in an attempt to make the Support Vector Machines (SVM),
More informationBayesian Machine Learning (ML): Modeling And Inference in Big Data. Zhuhua Cai Google, Rice University caizhua@gmail.com
Bayesian Machine Learning (ML): Modeling And Inference in Big Data Zhuhua Cai Google Rice University caizhua@gmail.com 1 Syllabus Bayesian ML Concepts (Today) Bayesian ML on MapReduce (Next morning) Bayesian
More informationEnvironmental Remote Sensing GEOG 2021
Environmental Remote Sensing GEOG 2021 Lecture 4 Image classification 2 Purpose categorising data data abstraction / simplification data interpretation mapping for land cover mapping use land cover class
More informationHow To Optimize Online Video Ads
Proceedings of the 2011 IEEE Systems and Information Engineering Design Symposium, University of Virginia, Charlottesville, VA, USA, April 29, 2011 FridayPM1Modeling.2 Bidding Strategies Optimization for
More informationSTA 4273H: Statistical Machine Learning
STA 4273H: Statistical Machine Learning Russ Salakhutdinov Department of Statistics! rsalakhu@utstat.toronto.edu! http://www.cs.toronto.edu/~rsalakhu/ Lecture 6 Three Approaches to Classification Construct
More informationEmotion Detection from Speech
Emotion Detection from Speech 1. Introduction Although emotion detection from speech is a relatively new field of research, it has many potential applications. In human-computer or human-human interaction
More informationMachine Learning Capacity and Performance Analysis and R
Machine Learning and R May 3, 11 30 25 15 10 5 25 15 10 5 30 25 15 10 5 0 2 4 6 8 101214161822 0 2 4 6 8 101214161822 0 2 4 6 8 101214161822 100 80 60 40 100 80 60 40 100 80 60 40 30 25 15 10 5 25 15 10
More informationLearning is a very general term denoting the way in which agents:
What is learning? Learning is a very general term denoting the way in which agents: Acquire and organize knowledge (by building, modifying and organizing internal representations of some external reality);
More informationK-Means Clustering Tutorial
K-Means Clustering Tutorial By Kardi Teknomo,PhD Preferable reference for this tutorial is Teknomo, Kardi. K-Means Clustering Tutorials. http:\\people.revoledu.com\kardi\ tutorial\kmean\ Last Update: July
More information1. Classification problems
Neural and Evolutionary Computing. Lab 1: Classification problems Machine Learning test data repository Weka data mining platform Introduction Scilab 1. Classification problems The main aim of a classification
More information1816 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 15, NO. 7, JULY 2006. Principal Components Null Space Analysis for Image and Video Classification
1816 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 15, NO. 7, JULY 2006 Principal Components Null Space Analysis for Image and Video Classification Namrata Vaswani, Member, IEEE, and Rama Chellappa, Fellow,
More informationSPECIAL PERTURBATIONS UNCORRELATED TRACK PROCESSING
AAS 07-228 SPECIAL PERTURBATIONS UNCORRELATED TRACK PROCESSING INTRODUCTION James G. Miller * Two historical uncorrelated track (UCT) processing approaches have been employed using general perturbations
More informationDesign of an FX trading system using Adaptive Reinforcement Learning
University Finance Seminar 17 March 2006 Design of an FX trading system using Adaptive Reinforcement Learning M A H Dempster Centre for Financial Research Judge Institute of Management University of &
More information10-601. Machine Learning. http://www.cs.cmu.edu/afs/cs/academic/class/10601-f10/index.html
10-601 Machine Learning http://www.cs.cmu.edu/afs/cs/academic/class/10601-f10/index.html Course data All up-to-date info is on the course web page: http://www.cs.cmu.edu/afs/cs/academic/class/10601-f10/index.html
More informationA Simple Introduction to Support Vector Machines
A Simple Introduction to Support Vector Machines Martin Law Lecture for CSE 802 Department of Computer Science and Engineering Michigan State University Outline A brief history of SVM Large-margin linear
More informationMachine Learning Logistic Regression
Machine Learning Logistic Regression Jeff Howbert Introduction to Machine Learning Winter 2012 1 Logistic regression Name is somewhat misleading. Really a technique for classification, not regression.
More informationLecture/Recitation Topic SMA 5303 L1 Sampling and statistical distributions
SMA 50: Statistical Learning and Data Mining in Bioinformatics (also listed as 5.077: Statistical Learning and Data Mining ()) Spring Term (Feb May 200) Faculty: Professor Roy Welsch Wed 0 Feb 7:00-8:0
More informationAnalyzing the Procurement Process in a Simplified Version of the TAC SCM Game
Analyzing the Procurement Process in a Simplified Version of the TAC SCM Game Hosna Jabbari December 9, 2005 Abstract The TAC supply chain management game presents a simulated market environment with automated
More informationLaboratory work in AI: First steps in Poker Playing Agents and Opponent Modeling
Laboratory work in AI: First steps in Poker Playing Agents and Opponent Modeling Avram Golbert 01574669 agolbert@gmail.com Abstract: While Artificial Intelligence research has shown great success in deterministic
More informationSIMPLIFIED PERFORMANCE MODEL FOR HYBRID WIND DIESEL SYSTEMS. J. F. MANWELL, J. G. McGOWAN and U. ABDULWAHID
SIMPLIFIED PERFORMANCE MODEL FOR HYBRID WIND DIESEL SYSTEMS J. F. MANWELL, J. G. McGOWAN and U. ABDULWAHID Renewable Energy Laboratory Department of Mechanical and Industrial Engineering University of
More informationPLAANN as a Classification Tool for Customer Intelligence in Banking
PLAANN as a Classification Tool for Customer Intelligence in Banking EUNITE World Competition in domain of Intelligent Technologies The Research Report Ireneusz Czarnowski and Piotr Jedrzejowicz Department
More informationTrading Tutorial. Microstructure 2 Liquidity. Rotman School of Management http://rit.rotman.utoronto.ca Page 1 of 9
Microstructure 2 Tutorial Introduction Welcome to Rotman Interactive Trader (RIT). This document begins assuming that you have installed and run the Rotman Interactive Trader client application (RIT Client).
More information