Abstract This study consists introductory survey and three essays where investors' trading responses to interim earnings announcements are studied using Finnish data. The essays are individual papers, but their topics are closely connected since they address the trading response from different angles. The essays progress from an aggregated to a more detailed examination. The first essay was conducted on daily data, whereas the second and third consist of intraday trading data. In all three essays information asymmetry is assumed to affect trading behavior around interim earnings announcements. The first article contains empirical findings regarding the effect of interim earnings announcements on investors' trading policy using Finnish data. The aim of the paper is to investigate empirically the role of pre-disclosure information asymmetry and the information content in explaining volume responses to interim earnings announcements. Evidence is provided that the trading volume response is positively associated with the information content and to some extent with the level of pre-disclosure information asymmetry. The results are in line with the theoretical trading volume proposition. However, the significance levels are lower than in similar US studies and the association between positive and negative news is slightly asymmetric. The second article finds evidence from the Helsinki Stock Exchange that the widely documented U-shape pattern in trading activity - namely heavy trading in the beginning and at the end of the trading day and relatively light trading in the middle of the day - is affected by an anticipated information event (i.e. interim earnings announcement). Before the announcement day, trading is more concentrated at the close. This is consistent with investors' heterogeneous willingness to bear expected overnight risk, which is especially prevalent before an announcement. Moreover, a slight increase on the open is evident after the announcement day. Evidence is also provided that the change in intraday trading behavior is associated with announcement-related factors, such as the range of analysts' earnings forecasts, the magnitude of unexpected earnings and firm size. Furthermore, this association is evident to some extent during the transition between trading and non-trading regimes. The third study examines whether the permanent price effects of individual trades are greater before or after an interim earnings announcement on the Helsinki Stock Exchange. If the permanent price effects are greater before the announcement this would suggest that investors believe that some traders are better informed before the interim earnings announcement than after. Using permanent price effects as a measure of price adjustment for private information, tests were performed to see whether price adjustments are greater in pre-announcement periods than in post-announcement periods. The results, based on interim earnings releases for the period 1993 to 1997 by HSE-listed firms, suggest that large trades do indeed produce greater permanent price effects before an announcement than after it. This suggests that large trades associated with price changes (especially uptick trades) before an announcement send a stronger signal to other investors than similar trades after the announcement. For small trades the results were insignificant.

This study assesses different machine learning and statistical methods to perform short-term point electricity price forecasting on the maximum buying and minimum selling Intraday (ID) market prices for each hour. The study begins with a primer and background on the current state of the electricity markets and why the need to trade on an ID market is growing.  The study examines different time-series forecasting methods using available exogenous electricity market data, such as the Day-ahead (DA) market price data and the ID prices. The models are evaluated on a set of error metrics, and for comparison, is a baseline constructed by using the DA price for the same hour as the forecast for the targets. The models evaluated are a Deep Neural Nets (DNN) model, an Autoregressive (AR) model and a XGBoost model. Further, a data scaling and transformation method, referred to as the Median-normalised asinh Transform (asinh1), improves the performance of all the models except the baseline, compared to Standardisation scaling (StdSc). The regularised AR model performed best, with the lowest overall scores on the metrics. However, the DNN model can best capture outlier patterns in the minimum selling ID prices. Throughout the study, it turns out that the buying price patterns and outliers are harder to forecast than the selling prices. This study aims to provide insights into the performance of different models and generally contribute to decreasing the knowledge gap between ID price forecasting and other electricity entities forecasting.

This thesis develops and applies methodological approaches for the analysis of intraday markets for electricity which are organised as continuous double auctions. The focus is to improve the understanding of how balancing forecast errors from weather-dependent renewable energy sources influences the outcomes of continuous intraday markets. This is important as it helps to assess how large amounts of renewable capacity can be utilised cost-efficiently and without stressing security of supply. In a first step, the thesis proposes a (non-mathematical) model of a continuous intraday market to show how the direction of the forecast error determines transactions between market participants, how these transactions relate to the formation of prices, and how the market integration of renewables can be improved. In a second step, the thesis provides a foundation for quantitative market analyses by modelling price-setting decisions for power generators and electricity demanders. This makes it possible to show that information on market participants' technical characteristics enables informed predictions of their market behaviour. In a third step, the thesis presents a computer simulation of a continuous intraday market. Implementing the simulation approach for the German power system allows calculation of the costs associated with the uncertain feed-in from renewables.

[pt] Há diferentes tipos de investidores que compõem o mercado financeiro e produzem oportunidades de mercado em diferentes escalas de tempo. Isto evidencia uma estrutura heterogênea de mercado. Nesta tese conjecturamos que podem haver oportunidades mais preditivas do que outras, o que motiva a investigação e a construção de estratégias multirresolução. Para estratégias multirresolução há abordagens que utilizam a decomposição de séries temporais para a operação em resoluções distintas ou propostas para a construção de conjuntos de dados de acordo com decisões de negociação multirresolução. As demais estratégias, em sua maioria, são de resolução única. Nesta tese, abordamos dois problemas, maximização de retorno acumulado e maximização de retorno acumulado com o risco controlado, e propomos uma abordagem computacionalmente eficiente para a construção de estratégias multirresolução, a partir da resolução do problema de Agendamento de Intervalos Ponderados. Nossa metodologia consiste em dividir o dia de mercado em intervalos, especializar traders por intervalo e associar um prêmio a cada trader. Para o problema de maximização de retorno acumulado, o prêmio de cada trader corresponde ao retorno acumulado entre dias para o intervalo de operação associado. Para o problema de maximização de retorno acumulado com controle do risco, o prêmio de cada trader corresponde ao retorno acumulado dividido pelo risco para o intervalo de operação associado. Diferentemente do problema anterior, empregamos um conjunto de traders por intervalo e utilizamos o método de Média-Variância, de Markowitz, para encontrar pesos ótimos para conjunto de traders de forma a controlar o risco. Conjecturamos aqui que o controle do risco por intervalo acarreta no controle do risco global da estratégia para o dia. Para a sinalização das ordens de compra e venda, nossos traders utilizam detectores de oportunidades. Estes detectores utilizam algoritmos de Aprendizado de Máquina que processam informações de indicadores de análise técnica e dados de preço e volume. Realizamos experimentos para dez ativos de maior liquidez da BMF&Bovespa para um período de um ano. Nossa estratégia de Composição de um Time de Traders (CTT) apresenta 0, 24 por cento de lucro médio diário e 77, 24 por cento de lucro anual, superando em 300 por cento e 380 por cento, respectivamente, uma estratégia de resolução única. Para os custos adotados, a estratégia CTT é viável a partir de 50.000,00 dólares. Para o problema de maximização do retorno acumulado com risco controlado, a estratégia de Composição de Carteiras por Intervalos (CCI) apresenta em média 0, 179 por cento de lucro diário e 55, 85 por cento de lucro anual, superando o método de Média-Variância de Markowitz. Para os custos adotados, a estratégia CCI é viável a partir de 2.000.000,00 dólares. As principais contribuições desta tese são: abordagem por Agendamentos de Intervalos Ponderados para a construção de estratégias e o emprego do modelo de Média-Variância para compor uma carteira de traders ao invés da tradicional abordagem por ativos. / [en] There are different types of investors who make up the financial market and produce market opportunities at different time scales. This indicates a heterogeneous market structure. In this thesis, we conjecture that may have more predictive opportunities than others, what motivates research and construction of we denominate multirresolution optimal strategies. For multirresolution strategies there are time series decomposition approaches for operating at different resolutions or proposals for dataset construction according to multirresolution trading optimal decisions. The other approaches, are single resolution. Thus, we address two problems, maximizing cumulative returns and maximizing cumulative returns with risk control. Here, we propose solving the Weighted Interval Scheduling problem to build multirresolution strategies. Our methodology consists of dividing the market day into time intervals, specialize traders by interval and associate a prize to each trader. For the cumulative return maximization problem, the prize corresponds to cumulative returns between days for the associated trader operation interval. For the cumulative return maximization problem with risk control each trader prize corresponds to cumulative return divided by risk with associated operation interval. In order to control the risk, we employ a set of traders by interval and apply the Markowitz Mean-Variance method to find optimal weight for set of traders. Here, we conjecture that controlling each interval risk leads to the overall risk control of the day. For signaling buy and sell orders, our traders use opportunity detectors. These detectors correspond to Machine Learning algorithms that process technical analysis indicators, price and volume data. We conducted experiments for ten of the most liquid BMF&Bovespa stocks to a one year span. Our Trading Team Composition strategy results indicates an average of 0.24 per cent daily profit and a 77.24 per cent anual profit, exceeding by 300 per cent and 380 per cent, respectively, a single resolution strategy. Regarding operational costs, CTT strategy is viable from 50,000 dollars. For the cumulative return maximization problem under risk control, our Portfolio Composition by Intervals strategy results indicates an average of 0.179 per cent daily profit and a 55.85 per cent anual profit, exceeding a Markowitz Mean- Variance method. Regarding operational costs, CCI strategy is viable from 2,000,000 dollars. Our main contributions are: the Weighted Interval Scheduling approach for building multirresolution strategies and a portfolio composition of traders instead of stocks performances.

This thesis develops and applies methodological approaches for the analysis of intraday markets for electricity which are organised as continuous double auctions. The focus is to improve the understanding of how balancing forecast errors from weather-dependent renewable energy sources influences the outcomes of continuous intraday markets. This is important as it helps to assess how large amounts of renewable capacity can be utilised cost-efficiently and without stressing security of supply. In a first step, the thesis proposes a (non-mathematical) model of a continuous intraday market to show how the direction of the forecast error determines transactions between market participants, how these transactions relate to the formation of prices, and how the market integration of renewables can be improved. In a second step, the thesis provides a foundation for quantitative market analyses by modelling price-setting decisions for power generators and electricity demanders. This makes it possible to show that information on market participants' technical characteristics enables informed predictions of their market behaviour. In a third step, the thesis presents a computer simulation of a continuous intraday market. Implementing the simulation approach for the German power system allows calculation of the costs associated with the uncertain feed-in from renewables.

Commonly, the day-ahead and intraday market on the electricity exchange are treated separately in academia. However, a model that forecasts the direction of the price spread between these two markets creates an opportunity for a market participant to leverage the price spread. In the neighbouring domain, electricity price forecasting, deep learning has proven to excel. Therefore, it is hypothesised that it will do so in directional price spread forecasting as well. A quantitative case study was performed to investigate how accurately a deep learning approach could be in directional electricity price spread forecasting. The case study was conducted on the Nordic electricity exchange Nord Pool in the SE3 region. The deep learning approach was compared with previously suggested machine learning models and a naive heuristic. The results show no statistical difference in error rate between the deep learning model and the machine learning model or naive heuristic. The results suggest that deep learning might not be a suitable approach to the task or that the implementation did not fully exhaust the potential of deep learning. / Vanligtvis behandlas marknaden för day-ahead och intraday på elbörsen separat i den akademiska litteraturen. En modell som prognostiserar riktningen för prisskillnaden mellan dessa två marknader skapar dock en möjlighet för en marknadsaktör att utnyttja prisskillnaden. I grannområdet elprisprognoser har djupinlärning visat sig överträffa andra typer av modeller. Därför antas det att djupinlärning även kommer göra det i riktade prisskillnadsprognoser. En kvantitativ fallstudie utfördes för att undersöka hur precis en djupinlärningsmetod kan vara i prognos för riktad elprisskillnad. Fallstudien genomfördes på den nordiska elbörsen Nord Pool i SE3-regionen. Djupinlärningsmetoden jämfördes med tidigare föreslagna maskininlärningsmodeller och en naiv heuristik. Resultaten visar ingen statistisk skillnad i fel-andel mellan djupinlärningsmodellen och maskininlärningsmodellen eller naiv heuristik. Resultaten antyder att djupinlärning kanske inte är ett lämpligt tillvägagångssätt för uppgiften eller att implementeringen inte helt utnyttjar potentialen för djupinlärning.

This study examines the comparative magnitude of disturbances in intraday data for exchange traded foreign exchange (FX) options. An in-depth time series analysis on the frequency and extent of discrepancies in the disturbances is conducted. The purpose of this study is twofold. First, using intraday data and trading volume, this study attempts to determine whether both put-call parity and lower boundary conditions consistently hold for exchange traded options written on U.S. dollar denominated options on the Euro trading on the Philadelphia Stock Exchange (PHLX). Second, this study attempts to investigate the magnitude of any discrepancies that may exist due to a temporary cessation of either put-call parity or lower boundary conditions. Intraday (tick-by-tick) bid prices, ask prices, and trading volume on U.S. dollar denominated European style call options and put options on the Euro are obtained. Option data is collected through a Structured Query Language (SQL) request from the Bloomberg database. Corresponding tick-by-tick spot rates for the underlying exchange rate are obtained for the same time period. Tick-by-tick 3-month Treasury bill rates are obtained to for use as the relevant risk-free interest rate. The primary data set spans an approximate one month period from 11/1/2011 to 12/6/2011. Call and option pricing data for near-the-money exercise prices are obtained for options expiring in December 2011, January 2012, February 2012, March 2012, June 2012, and September 2012. A total of 7,212 ticks (minutes) are analyzed for the conversion strategy and 7,209 ticks are analyzed for the reversal strategy. The data is structured into an unbalanced panel data set (cross-sectional time series data) using put-call pairs as the cross sectional units and ticks as the time-series unit. To test the efficiency of the foreign exchange options market, lower boundary and put-call parity conditions were tested on tick-by-tick currency option data. Analysis shows that lower boundary conditions hold for the overwhelming majority of options, with less than 0.0001% of violations for the observed options. A more detailed econometric analysis was prepared to test the put-call parity condition for currency options. A fixed effects model specification is used to describe the put-call parity relationship. Based on the analysis, it is possible to obtain arbitrage profits in the short run through the use of either a conversion or reversal strategy even after accounting for transaction costs. Taking the first differences of the variables resulted in a model with stationary variables and statistically significant estimators. The inclusion of dummy variables for moneyness did not add significant explanatory power to the deterministic put-call parity relationship. For both first differences of conversion and reversal strategies, the large t-statistics for the slope coefficients and intercept terms indicate a rejection of the null hypothesis, H0: λ0 = 0 and λ1 = 1 after adjusting for standard error. This implies that once transaction costs are adjusted for, put-call parity does not hold. However, the intercept term is only very slightly negative, and the intercept term is only slightly less than one in both cases. This implies that when put-call parity is violated, arbitrage profit should be relatively small.

This thesis comprises four papers concerning trade durations and limit order book information. Paper [1], [2] and [4] study trader durations, e.g., the time between stock transactions in intra-day data. Paper [3] focus on the information content in the limit order book concerning future price movements in stock transaction data. Paper [1] considers conditional duration models in which durations are in continuous time but measured in grouped or discretized form. This feature of recorded durations in combination with a frequently traded stock is expected to negatively influence the performance of conventional estimators for intraday duration models. A few estimators that account for the discreteness are discussed and compared in a Monte Carlo experiment. An EM-algorithm accounting for the discrete data performs better than those which do not. Empirically, the incorporation of level variables for past trading is rejected in favour of change variables. This enables an interpretation in terms of news effects. No evidence of asymmetric responses to news about prices and spreads is found. Paper [2] considers an extension of the univariate autoregressive conditional duration model to which durations from a second stock are added. The model is empirically used to study duration dependence in four traded stocks, Nordea, Föreningssparbanken, Handelsbanken and SEB A on the Stockholm Stock Exchange. The stocks are all active in the banking sector. It is found that including durations from a second stock may add explanatory power to the univariate model. We also find that spread changes have significant effect for all series. Paper [3] empirically tests whether an open limit order book contains information about future short-run stock price movements. To account for the discrete nature of price changes, the integer-valued autoregressive model of order one is utilized. A model transformation has an advantage over conventional count data approaches since it handles negative integer-valued price changes. The empirical results reveal that measures capturing offered quantities of a share at the best bid- and ask-price reveal more information about future short-run price movements than measures capturing the quantities offered at prices below and above. Imbalance and changes in offered quantities at prices below and above the best bid- and askprice do, however, have a small and significant effect on future price changes. The results also indicate that the value of order book information is short-term. Paper [4] This paper studies the impact of news announcements on trade durations in stocks on the Stockholm Stock Exchange. The news are categorized into four groups and the impact on the time between transactions is studied. Times before, during and after the news release are considered. Econometrically, the impact is studied within an autoregressive conditional duration model using intradaily data for six stocks. The empirical results reveal that news reduces the duration lengths before, during and after news releases as expected by the theoretical litterature on durations and information flow.

As science advances, computer technologies are developing rapidly in the past decades. The previous way of traders¡¦ yelling for orders in the house of exchange has been replaced by the Internet and computers. The trading modes of institutional investors are transforming gradually, particularly the radical changes in the US stock market for the past 5 years. The transaction volume from high frequency trading and algorithmic trading is growing dramatically per year, accounting for at least 70% in the U.S. market. And many researchers find these trading methods based on the computer programs good in increasing liquidity, reducing volatility and facilitating price discovery. By using intraday data of Taiwan stock market in 2008 to conduct empirical research, this study intends to analyze the effect of this trend on the TW stock market. Empirical results found that the greater the market capitalization, liquidity, stock volatility are, the higher the proportion of algorithmic trading will be, but which only exists in foreign institutional investors. On the other hand, the increase of the proportion of algorithmic trading can improve liquidity, meanwhile raise the volatility. The conclusion remains unchanged when applied to control the effect of financial tsunami. That means algorithmic trader¡¦s behaviors are not always positive. This result could be related to the special transaction mechanism or lower competition of algorithmic trading in Taiwan. As to trading strategy, the result found that foreign institutional investors focus on momentum strategies, whereas particular dealers act for the sake of index arbitrage or hedge. In summary, the algorithmic trader¡¦s transaction bears positive (liquidity) and negative (volatility) impact on the market at the same time. For individual investors, algorithmic trading¡¦s momentum strategy could appeal to them, but they may not make a profit from these trades, because this strategy could merely want to pull price higher and sell stock or the opposite. About regulators, algorithmic traders¡¦ behavior should be regulated partly; regulatory authorities might also consider adding the circuit mechanism similar to South Koreas¡¦, especially on the program trading. Keywords: algorithmic trading, high frequency trading, intraday, strategy, liquidity, volatility, market quality

none