Assessing cartel detection and damages in simulated markets : a comparative study of econometric and machine learning approaches

Date
2024-03
Journal Title
Journal ISSN
Volume Title
Publisher
Stellenbosch : Stellenbosch University
Abstract
ENGLISH SUMMARY: Collusion among firms, with the intent to artificially elevate prices, has far-reaching implications for market competition and consumer welfare. This thesis explores the detection of structural breaks in simulated price data under collusion, and their subsequent impact on damage estimation. This contribution is significant in the field of competition economics as it allows for consideration of the implications for econometric methods aimed at identifying and measuring collusive effects in the age of machine learning alternatives. A combination of econometric and machine learning approaches, including Lasso regression, random forest regression and classification, logistic regression, and Bai-Perron structural break testing are rigorously examined against four distinct data generating processes simulated to mimic the behaviours of cartels observed in the market. These include a deterministic switch data generating process, a recurrent switch data generating process, a phased switch data generating process, and a Markov-switching data generating process. The study reveals that the Lasso model consistently outperforms the other methods in estimating structural breaks, demonstrating superior performance in identifying cartel and competitive pricing behaviours across the different linear data generating processes. Conversely, the Bai-Perron test exhibits the poorest performance, particularly in Phase and Markov-switching transitions, highlighting its limitations in capturing nuanced structural changes. Furthermore, damage estimation was performed using dummy variables generated by each of the models. All of the empirical models perform relatively well in capturing damages, with the exception of the Bai-Perron model when applied to the phase and Markov-switching data generating processes, further emphasising its limited utility in detecting nuanced switching mechanisms in pricing behaviour. To enhance the analysis, damage estimation was alternatively conducted by predicting movements in the price variable for the Lasso and random forest models. These modifications revealed slight discrepancies in damage predictions, with the Lasso model overpredicting and the random forest model underpredicting damages. Nevertheless, both models remain highly accurate in capturing the economic impact of structural changes in competitive pricing. This research contributes to the field of competition economics by providing a comprehensive analysis of structural break detection and damage estimation methodologies, ultimately demonstrating the practical advantages of the Lasso regression model when applied to linear pricing models. These findings offer valuable insights for policymakers and analysts seeking to better understand and address changes in competitive market dynamics.
AFRIKAANSE OPSOMMING: Samespanning tussen firmas, met die doel om pryse kunsmatig te verhoog, het verreikende implikasies vir markmededinging en verbruikerswelsyn. Hierdie tesis ondersoek die opsporing van strukturele onderbrekings in kompetisie-ekonomiese data en die daaropvolgende impak daarvan op skadeberaming. Hierdie bydrae is betekenisvol in die veld van mededingingsekonomie, aangesien dit voorsiening maak vir oorweging van die implikasies vir ekonometriese metodes wat daarop gemik is om samespanningseffekte in die era van masjienleer-alternatiewe te identifiseer en te meet. n’ Kombinasie van ekonometriese en masjienleer benaderings, insluitend Lasso-regressie, Ewekansige Woud-regressie en klassifikasie, Logistiese regressie en Bai-Perron-strukturele breuk toetsing is breedvoerig ondersoek teen vier verskillende data genererende prosesse wat die gedrag van kartelle in verskeie market simuleer. Hierdie sluit in n’ deterministiese, herhalende, gefaseerde en Markov skakeldatagenereringsprosese. Die studie onthul dat die Lasso-model konsekwent beter presteer as die ander metodes in die skatting van strukturele onderbrekings, wat uitstekende prestasie demonstreer in die identifisering van kartel- en mededingende prysgedrag oor die verskillende lineere datagenererende prosesse. Omgekeerd toon die Bai-Perron-toets die swakste prestasie, veral in fase- en Markov-wissel-oorgange, wat die beperkings daarvan beklemtoon om nuanse in strukturele veranderinge vas te le. Verder is skadeskatting gebou deur die gebruik van skyn-veranderlikes wat deur elk van die modelle gegenereer is. In hierdie aspek presteer alle modelle relatief goed in die vaslegging van skade, met die uitsondering van die Bai-Perron-model wanneer dit toegepas word op die Fase- en Markov-skakeling data genererende prosesse, wat die beperkte nut daarvan verder beklemtoon in die opsporing van genuanseerde skakelmeganismes. Om die ontleding te verbeter, is skadeberaming gedoen deur die prysveranderlike vir die Lassoen random forest-modelle te teiken. Hierdie veranderinge het klein teenstrydighede in skadevoorspellings gebring, met die Lasso-model wat oorvoorspel en die random forest-model skade ondervoorspel. Nietemin, beide modelle bly hoogs akkuraat in die vaslegging van die ekonomiese impak van strukturele veranderinge in mededingende pryse. Hierdie navorsing dra by tot die veld van mededingingsekonomie deur 'n omvattende ontleding van strukturele breukopsporing en skadeberamingsmetodologiee te verskaf, wat uiteindelik die praktiese voordele van die Lasso-regressiemodel demonstreer. Hierdie bevindinge bied waardevolle insigte vir beleidmakers en ontleders wat probeer om veranderinge in mededingende markdinamika beter te verstaan en aan te spreek.
Description
Thesis (MCom)--Stellenbosch University, 2024.
Keywords
Citation