In this work, we propose a regressional framework, built on the work ”Deep Order Flow Imbalance: Extracting Alpha at Multiple Horizons from the Limit Order Book” by Kolm, et al. (2023), for predicting short term returns of odds on binary betting exchange markets. Using the framework, we apply five different deep learning models that leverage order book data from tennis betting exchanges during the calendar month of July 2023 with the purpose of examining the predictive capabilities of deep learning models in this setting. We train each model on either raw limit order book states or order flow. The models predict the returns of the best available odds returns on five different short term time horizons on the four order book sides, back and lay for each of the two players in a given tennis match. Applying windowing, for each vector prediction we use the 100 latest market messages consisting of 81 features (odds and volumes per the ten first levels in the order book and time delta between market messages) in the case of the raw limit order book state and 41 features (order book flow per the ten first levels in the order book and time delta between market messages) in the case of the order book flow. All code is written in Python and run on Google Colab, leveraging cloud computing, off-the-shelf models and popular libraries, TensorFlow and Keras, for data processing and pipelining, model implementation, training and testing. The models are evaluated relative to a benchmark in the form of a naive predictor based on the average odds returns on the training set. The models do not converge towards an optimal parameter composition duringtraining, indicating low predictive capabilities of the input data. Despite this, we generally find all models to outperform the benchmark on the lay order book sides and while some perform better than others, we see similar relative performance distributions within each model across horizon-order book side combinations. To enhance discussion and suggest the direction of future research we examine relationships between key game characteristics such asthe variation of odds returns and the accuracy of predictions on a given market.

En utilisant des données haute-fréquence inédites, cette thèse étudie trois types de regroupements (“clusters”) présents dans le marché des changes: la concentration d'ordres sur certains prix, la concentration des transactions dans le temps et l'existence de groupes d'investisseurs prenant les mêmes décisions. Nous commençons par étudier les propriétés statistiques du carnet d'ordres EBS pour les paires de devises EUR/USD et USD/JPY et l'impact d'une réduction de la taille du tick sur sa dynamique. Une grande part des ordres limites est encore placée sur les anciens prix autorisés, entraînant l'apparition de prix-barrières, où figurent les meilleures limites la plupart du temps. Cet effet de congestion se retrouve dans la forme moyenne du carnet où des pics sont présents aux distances entières. Nous montrons que cette concentration des prix est causée par les traders manuels qui se refusent d’utiliser la nouvelle résolution de prix. Les traders automatiques prennent facilement la priorité, en postant des ordres limites un tick devant les pics de volume.Nous soulevons ensuite la question de l'aptitude des processus de Hawkes à rendre compte de la dynamique du marché. Nous analysons la précision de tels processus à mesure que l'intervalle de calibration est augmenté. Différent noyaux construits à partir de sommes d'exponentielles sont systématiquement comparés. Le marché FX qui ne ferme jamais est particulièrement adapté pour notre but, car il permet d’éviter les complications dues à la fermeture nocturne des marchés actions. Nous trouvons que la modélisation est valide selon les trois tests statistiques, si un noyau à deux exponentielles est utilisé pour fitter une heure, et deux ou trois pour une journée complète. Sur de plus longues périodes la modélisation est systématiquement rejetée par les tests à cause de la non-stationnarité du processus endogène. Les échelles de temps d'auto-excitation estimées sont relativement courtes et le facteur d'endogénéité est élevé mais sous-critique autour de 0.8. La majorité des modèles à agents suppose implicitement que les agents interagissent à travers du prix des actifs et des volumes échangés. Certains utilisent explicitement un réseau d'interaction entre traders, sur lequel des rumeurs se propagent, d'autres, un réseau qui représente des groupes prenant des décisions communes. Contrairement à d'autres types de données, de tels réseaux, s'ils existent, sont nécessairement implicites, ce qui rend leur détection compliquée. Nous étudions les transactions des clients de deux fournisseur de liquidités sur plusieurs années. En supposant que les liens entre agents sont déterminés par la synchronisation de leur activité ou inactivité, nous montrons que des réseaux d'interactions existent. De plus, nous trouvons que l'activité de certains agents entraîne systématiquement l’activité d'autres agents, définissant ainsi des relations de type “lead-lag” entre les agents. Cela implique que le flux des clients est prévisible, ce que nous vérifions à l'aide d'une méthode sophistiquée d'apprentissage statistique. / The aim of this thesis is to study three types of clustering in foreign exchange markets, namely in price, trades arrivals and investors decisions. We investigate the statistical properties of the EBS order book for the EUR/USD and USD/JPY currency pairs and the impact of a ten-fold tick size reduction on its dynamics. A large fraction of limit orders are still placed right at or halfway between the old allowed prices. This generates price barriers where the best quotes lie for much of the time, which causes the emergence of distinct peaks in the average shape of the book at round distances. Furthermore, we argue that this clustering is mainly due to manual traders who remained set to the old price resolution. Automatic traders easily take price priority by submitting limit orders one tick ahead of clusters, as shown by the prominence of buy (sell) limit orders posted with rightmost digit one (nine).The clustering of trades arrivals is well-known in financial markets and Hawkes processes are particularly suited to describe this phenomenon. We raise the question of what part of market dynamics Hawkes processes are able to account for exactly. We document the accuracy of such processes as one varies the time interval of calibration and compare the performance of various types of kernels made up of sums of exponentials. Because of their around-the-clock opening times, FX markets are ideally suited to our aim as they allow us to avoid the complications of the long daily overnight closures of equity markets. One can achieve statistical significance according to three simultaneous tests provided that one uses kernels with two exponentials for fitting an hour at a time, and two or three exponentials for full days, while longer periods could not be fitted within statistical satisfaction because of the non-stationarity of the endogenous process. Fitted timescales are relatively short and endogeneity factor is high but sub-critical at about 0.8.Most agent-based models of financial markets implicitly assume that the agents interact through asset prices and exchanged volumes. Some of them add an explicit trader-trader interaction network on which rumors propagate or that encode groups that take common decisions. Contrarily to other types of data, such networks, if they exist, are necessarily implicit, which makes their determination a more challenging task. We analyze transaction data of all the clients of two liquidity providers, encompassing several years of trading. By assuming that the links between agents are determined by systematic simultaneous activity or inactivity, we show that interaction networks do exist. In addition, we find that the (in)activity of some agents systematically triggers the (in)activity of other traders, defining lead-lag relationships between the agents. This implies that the global investment flux is predictable, which we check by using sophisticated machine learning methods.

Submitted by Raphael Maaz (rfmaaz@gmail.com) on 2018-08-17T04:37:24Z No. of bitstreams: 1 Dissertation_20180817_RFM.pdf: 420668 bytes, checksum: 15c23fd991dfb5a7d5edff23f64e01b7 (MD5) / Rejected by Thais Oliveira (thais.oliveira@fgv.br), reason: Raphael, boa tarde! Para que possamos aprovar seu trabalho, a única alteração que deve ser feira é retirar o acento do "GETULIO" (nome da Escola). Por gentileza, alterar e submeter novamente. Qualquer dúvida, entre em contato. Thais Oliveira mestradoprofissional@fgv.br/ 3799-7764 on 2018-08-20T20:41:55Z (GMT) / Submitted by Raphael Maaz (rfmaaz@gmail.com) on 2018-08-21T03:59:50Z No. of bitstreams: 1 Dissertation_20180821_RFM.pdf: 420299 bytes, checksum: 16fbfa33cab672015314b8f39f931c77 (MD5) / Approved for entry into archive by Thais Oliveira (thais.oliveira@fgv.br) on 2018-08-21T23:50:47Z (GMT) No. of bitstreams: 1 Dissertation_20180821_RFM.pdf: 420299 bytes, checksum: 16fbfa33cab672015314b8f39f931c77 (MD5) / Approved for entry into archive by Suzane Guimarães (suzane.guimaraes@fgv.br) on 2018-08-22T13:12:56Z (GMT) No. of bitstreams: 1 Dissertation_20180821_RFM.pdf: 420299 bytes, checksum: 16fbfa33cab672015314b8f39f931c77 (MD5) / Made available in DSpace on 2018-08-22T13:12:56Z (GMT). No. of bitstreams: 1 Dissertation_20180821_RFM.pdf: 420299 bytes, checksum: 16fbfa33cab672015314b8f39f931c77 (MD5) Previous issue date: 2018-08-10 / O presente trabalho apresenta um estudo sobre catacterístitcas importantes da utilização de estratégias de negociações algorítmicas de alta frequência (HFT) por investidores que realizam operações com contratos futuros de taxa de câmbio em reais por dólar comercial no mercado brasileiro, tais como o nível de atividade de operadores que empregam este tipo de recurso e seus principais fatores de influência, a rentabilidade de participantes de instituições que adotam esta prática e a existência de padrões nas estratégias adotadas por negociadores algoritmicos. Dentre as conclusões obtidas, verificou-se que investidores que utilizam estratégias de HFT possuem uma participação de aproximadamente 70% no mercado de contratos futuros de dólar comercial. Além disso, o nível de participação é influenciado pela volatilidade de mercado e da taxa de câmbio futura, bem como o número de contratos em aberto, resultando em lucros diários de aproximadamente R$ 18 mil. Finalmente, verifica-se a existência de elevados níveis de correlação entre as estratégias utilizadas por negociadores algorítmicos e comportamento de manada, bem como a utilização de mecanismos de retroalimentação positiva. / This work presents a study about important characteristics of the adoption of high-frequency trading (HFT) strategies by investors that perform operations with USDBRL futures contracts on the Brazilian Foreign Exchange markets, such as the market share and the profits of the participants that resort to such strategies, as well as its main determinants, and the existence of patterns on the strategies adopted by high-frequency traders. It has been determined that investors that use HFT strategies possess 70% of market share on US dollar futures contracts and that its participation is influenced by markets and futures exchange rates volatility level, as well as its open interest, making an average profit of roughly 18 thousand Brazilian reals per day. Finally, it has been possible to determine the existence of high levels of correlation between the strategies adopted by high-frequency traders, as well as herding and the usage of positive feedback strategies.

The master thesis deals with proposal of automated trading system and its implementation in the Foreign exchange market environment. This system will be developed as investment model based on the analyzes performed and then tested on real data to achieve maximum stability and profit.

This thesis deals with a complex analysis and prediction of foreign exchange markets. It uses advanced artificial intelligence methods, namely neural networks and chaos theory. It introduces unconventional approaches and methods of each of these areas, compares them and uses on a real problem. The core of this thesis is a comparison of several prediction models based on completely different principles and underlying theories. The outcome is then a selection of the most appropriate prediction model called NAR + H. The model is evaluated according to several criteria, the pros and cons are discussed and approximate expected profitability and risk are calculated. All analytical, prediction and partial algorithms are implemented in Matlab development environment and form a unified library of all used functions and scripts. It also may be considered as a secondary main outcome of the thesis.