Stock trading, one of the most common economic activities in the world where the values of stocks change quickly over time. Some are able to turn great profits while others turn great losses on stock trading. Being able to predict changes could be of great help in maximising chances of profitability. In this report we want to evaluate the predictability of stock markets using Artificial Neural Network models, Adaptive Neuro-Fuzzy inference systems and Autoregressive-moving-average models. The markets used is Stockholm, Korea and Barcelona Stock Exchange. We are using two test scenarios, one which consists of incrementing the initial 25 days of training with 5 days until the end of the stock year, and the other one consisting of moving the 25 training days, 5 days until the stock year is over. Our results consists of showing how the predictions look like on the Stockholm market for all the methods and test scenarios and also show graphs of the error rate (percentage) of all methods and test cases in each of the markets. Also a table showing the average error of the methods and test cases, to be able to evaluate which one will perform the best. Our results shows that the ANFIS with at least 50 days of training will perform the best.

Brain computer interfaces (BCIs) enable communication between a brain and a computer, without the need for any motor actions. Electroencephalography (EEG) signals can be used as input for a BCI, but they need to go through a number of steps in the BCI to create useableoutput. One of the most critical steps is the classification algorithm,which is the step that is investigated in this report. A linear and anonlinear Support Vector Machine (SVM), together with a Linear Discriminant Analysis (LDA), are investigated in how well they can handlesession to session performance when classifying EEG data from three different recording sessions of three different test subjects. The results show that the average performance of the classifiers are in most cases similar, slightly above 60 %. The performance of the investigated algorithms differed depending on subject and session. The sometime slow performance of the classification algorithms may be due to the lowsignal-to-noise ratio in the EEG signals, or possibly even due to bad performance in producing recognizable EEG patterns by the test subjects.The conclusion drawn from the project is that data from different sessions can vary quite extensively, and in this project it was handledbest by the nonlinear SVM with RBF kernel, with the highest averageclassification accuracy. / Gränssnitt mellan hjärna och dator (BCI) möjliggör kommunikation utan behovet av någon motorisk rörelse. Elektroencefalografiska (EEG) signaler kan användas som indata till en BCI, men de behöver genomgåett antal steg i BCI:n för att göra om dem till användbar utdata. Ettav de mest kritiska stegen består av klassifikationsalgoritmen, vilket är det steg som undersöks i denna rapport. De algoritmer som undersöks i rapporten är en linjär och en ickelinjär Support Vector Machine(SVM), tillsammans med en Linear Discriminant Analysis (LDA), för att undersöka hur väl de kan hantera skillnader i EEG-datan från tre olika personer och sessioner. Resultaten visar att den genomsnittliga prestandan på klassifieringen i de flesta fallen är jämlik, strax ovanför60 %. Prestandan hos de olika klassifieringsalgoritmerna skiljer sig åt beroende på testperson och session. Den stundtals dåliga prestandan på klassifieringsalgoritmerna kan bero på den låga signal-till-brus kvotensom som finns hos EEG-signaler, men det kan möjligen även bero på dåliga prestationer i framkallandet av EEG-mönster hos testpersonerna. Slutsatsen som kan dras från projektet är att data från olika sessionerkan variera ganska mycket, och att i detta projektet hanterades detta bäst av den ickelinjära SVM:en med RBF-kärna som hade den högsta genomsnittliga klassifierings noggrannheten.

The stock market has a large impact on the economy of a nation, thisis why it is an interesting matter to see how stock market prediction canbe used and whether or not the predicted results are valid. This reportwill compare the prediction methods, the K Nearest Neighbour algorithmand the moving average formula using the closing prices of four Swedishequities that are based on the Stockholm stock exchange OMX. To geta proper familiarization into the background of stock markets and theutilized formulas, the report explains these theoretical concepts for thereader. A proper distribution of the results is given of the data with appropriatecharts and tables. Lastly a discussion explains the implicationsof the results and the conclusion that the K Nearest Neighbour algorithmproduced more accurate data when compared to the moving average formula. / Aktiemarknaden har en stor inverkan på en nations ekonomi, varför det är intressant att se om förutsägelser på aktiemarknaden kan användas samt om det förväntade resultatet är trovärdigt. Denna rapport kommer attjämföra slutkurser på fyra aktier frön Stockholmsbörsen med hjälp av K Närmaste Grannar algoritmen och det glidande medelvärdet formeln. För att ordentligt kunna sättas in i bakgrunden för aktiemarknaden och de valda formlerna, förklarar rapporten dessa villkor för läsaren. En lämplig fördelning av resultaten ges av det samlade datat med lämpliga diagramoch tabeller. Slutligen ges en diskussion som förklarar varför slutsatsen äratt K Närmaste Granne algoritmen ger ett mer exakt värde jämfört medden glidande medelvärde formeln.

The stock market plays a big role in our current nancial system and the uctuations onit are believed to depend on many dierent factors. One of the factors that are believedto be correlated to the stock market are macroeconomic variables, that is, variables thatindicate the status of the economical situation. Examples of such macroeconomic variablesare unemployment rate, loan interests and ination. Earlier attempts to predictthe stock market have been made by using process demanding methods such as arti-cial neural network. A multilayer perceptron is a self learning system that goes underthe category of an articial neural network. Such a network learns by being trainedon old data sets and has the capacity to identify relationships between dierent data.This method has been used in earlier studies to predict the stock market with goodresults. The problem statement of this report is to nd the optimal training interval fora multilayer perceptron on a day to day estimation of the Swedish OMXS30 index. Theinput to the algorithm consisted of 38 parameters, which in this case was a collectionof individual companies stock prices, foreign stock indexes, macroeconomic variables,previous and current values of the OMXS30 index. The results from the simulationsthat were executed on old stock data shows that 180 to 200 days of training yielded thebest results, where eight of nine periods over seven years (2007-2014) yielded prot. Theresults from the simulations during the periods with increasing index were sometimesbelow the index gain, but always with a prot. During periods of index decrease theresults were sometimes with a prot and sometimes non-prot. In the case of indexdecrease the result was always above the total index decrease. The conclusion is as theresults shows, that the optimal training interval is 180 to 200 days for the simulationsrun in the study of this report.1 / Aktiemarknaden spelar en stor roll i dagens finansiella system och fluktutionerna på börsen tros bero pa många orsaker. En av de saker som tros ha en koppling till börsen är makroekonomiska variabler, dvs sådana variabler som indikerar hur ekonomin mår. Exempel på makroekonomiska variabler ar arbetslöshet,       räntenivåer och i nation. Andra kopplingar som tros finnas till börsens utveckling är hur individuella aktier och utlandskabörser utvecklas. Tidigare försök har gjorts att forsöka forutsäga aktiemarknaden med hjalp av beräkningskrävande metoder, t. ex. Articiella neuron nät. En flerlagers perceptronar ett självlärande system som räknas som en typ av articiellt neuron nät. Nätverket lär sig genom att tränas pa gammal data och har formåagan att hitta samband mellan olika data. I tidigare studier har denna metod använts for att förutsäga aktiemarknaden med goda resultat. Problemformulering i denna rapport ar att ta reda på vilket det optimala träningsintervallet ar för en flerlagers perceptron för att, från en dag till en annan, förutsäga indexet på Stockholmsbörsen, OMXS30. Algoritmens indata bestod av totalt 38 parametrar som i detta fall var en samling av enskilda företagsaktievärden, utländska börsers index, makroekonomiska variabler, tidigare värden på OMXS30 samt det nuvarande värdet pa börsen. Resultaten från simulationerna som kördes pa gammal aktiedata visar att 180-200 dagar är det basta träningsintervallet daatta av nio stycken perioder över sju år (2007-2014) gick med vinst. Resultaten fransimulationerna under de perioder med stigande index blev i vissa fall under index, men alltid med vinst. I perioder med avtagande index sa blev resultaten i vissa fall vinstgivande och i andra fall inte vinstgivande, men i dessa fall alltid battre an den totalaindex nedgangen. Slutsatsen ar som resultaten visar att 180-200 dagar ar det optimala träningsintervallet for de simulationer som gjordes i undersökningen i denna rapport.2

The method used was to create an ensemble of classifiers, one for each time sample of asingle trial and thereafter using majority vote to decide the class of the trial. Classifiersused were support vector machines (SVMs) and linear discriminant analysis (LDA). Foreach subject data for 3 sessions were used, labeled A, B and C in chronological order.Session A and B were used as training sets, and session C as the test data. The result inthis study could not confirm the results of Herman et. al (2008) instead a slight positiveeffect of session A (average CA on session C 62%) compared to session B (average CAon session C 58%) could be seen, but in general there was no big difference in CA basedon the choice of training data (average CA on session C using training sets:A=62%, B=58%,  A&B=61%).Our results show that it is not always the case that training data recorded closer in time to the test data generate higher CA. Therefore we suggest that it could be a safer choice to use more than the latest session as training data. Still more studies are needed to confirmthat using more sessions for training really is better also on data where there is a bigger gap in performance between the latest and earlier sessions.

Trading indicators are frequently used among foreign exchange traders in attempts to predict future marketevents. Automated trading strategies can easily be implemented to act on such predictions. Motivated by acuriosity about whether the use of trading in dicators could be improved without actually changing the indicators themselves, this study was conducted in an attempt to investigate opportunities in enhancing strategy profits by restricting strategies from trading during periods deemed as unfavorable. However,conditionally restricting strategies’ trading capabilities by introducing thresholds for strategy activation did not show significant effects on the performance. By examining the accumulated strategy profits both with and without applied thresholds, it was derived that the general characteristics of the performance were withheld. Consequently, it can not be concluded that this study provides a reliable method of enhancing profits through applying restrictions to foreign exchange strategies. Nevertheless, the effects from applying thresholds to strategies, albeit not mainly profitable in this study, motivates further research on advantages from conditionally restricting foreign exchange strategies.

Schedule generation is a common real-world problem, that has been shown to behard to solve. In a scheduling algorithm, various constraints related to scheduling are the inputs and a schedule satisfying these constraints is the output. In this report, two algorithms for schedule generation are compared: Tabu Searchand a Genetic Algorithm. How well the algorithms perform for generating schedules from constraint input of dierent sizes is assessed, as well as how the performance of the algorithms is aected by varying parameters of the algorithm.The major conclusion drawn is that there is no major dierence between how well Tabu Search and the Genetic Algorithm scale when faced with a larger input size.

Forecasting short-term viewership ratings for on-demand video is crucial for the online advertisement market because advertisement sales is done ahead of time, and errors in forecasting means either loss of profit opportunities or having to compensate advertisers for not upholding agreements. These forecasts can be made using an uncomplicated Seasonal Averaging method, which produces forecasts for the coming weeks using averaged hourly values from previous weeks (where the forecast for next Sunday is the average of the actual value from the last three Sundays). In this thesis, an alternative approach using a neural network is implemented and benchmarked against the Seasonal Averaging method, using data from December–February from a major online videosite. The network utilizes a Multilayer Perceptron design with inputs corresponding to the seasonal patterns of the ratings data. It finds that while good forecasting performance can be reached even over very long horizons, weekly averages wins out when comparing standard forecasting error metrics, likely owing to the strong seasonal pattern. / Att förutsäga tittarsiffror för strömmande video är viktigt för reklamindustrin då försäljning av reklam sker innan den visats. Alltför stora fel i dessa förutsägelser leder till att annonsörer måste kompenseras för ej visade reklamsnuttar, alternativt att möjligheter till att sälja mer reklam går förlorade. Dessa förutsägelser kan göras genom att ta genomsnittet av tidigare veckors tittarsiffror och använda detta som förutsägelse för påföljande veckor (där tittarsiffran för söndag nästa vecka är lika med genomsnittet av de senaste tre söndagarna). I det här exjobbet undersöks möjligheten att använda ett neuronnätverk för att göra dessa förutsägelser istället, genom att jämföra resultaten från detta mot den nuvarande metoden på data från December till Februari. Neuronnätetär av typen Multilayer Perceptron och använder en design som är anpassat till de veckovisa mönster som data uppvisar. Undersökningen finner att trots goda förutsägelser från neuronnätverket når det inte samma träffsäkerhet (mätt med standardmått på förutsägelser) som den nu använda metoden, troligtvis på grund av det starka veckovisa mönstret som data uppvisar.

This study investigates the offline transcription capabilities of the Android Speech RecognitionEngine (ASRE). The need for this study comes from an original specification for an application tobe used to prevent illegal logging in the Amazon rainforest. Recognition of a set of spoken wordsand phrases was tested under controlled conditions in a recording studio using varying levels ofbackground noise. The results indicate that the ASRE can properly transcribe digits, parts of digits,and strings, but that cannot properly transcribe continuous text. The study finds that the ASREcould meet the needs of the original project specification and can be used to help prevent illegallogging in the Amazon rainforest. / Denna studie undersöker möjligheten till röstigenkänning i offlineläge hos Androidsröstigenkänningsmotor (ASRE). Studien behövs som en del i ett projekt för att producera enapplikation för att förhindra illegal avverkning i Amazonområdet. Tester av ASREs förmåga gjordesmed ord och fraser under kontrollerade förhållanden i en inspelningsstudio med varierade nivåer avbakgrundsljud. Resultaten visar att ASRE kan transkribera siffror, delar av siffror, och strängar, menatt den inte kan transkribera flytande text. Slutsatsen är att ASRE kan användas för att att uppfyllaoriginalspecifikationen och producera en applikation för att hjälpa till att förhindra illegalavverkning i Amazonområdet.Page 2

This thesis aimed to explore how sequential boolean satisability solvers can be used on the integer factorisation problem. The integer factorisation problem is believed to be hard and modern public key cryptography relies on that,note worthily SSL/TSL and SSH support the use of RSA. However, it is not proven that integer factorisation is hard and therefore it is of great importanceto explore dierent attack avenues. The modulus in RSA is a semiprime, e.g.an integer that is the product of two primes. Hence, in this thesis an empiricalstudy of the factorisation of semiprimes with a bit-length of up to 32 bits iscarried out. Randomly selected semiprimes are factorized through six dierent reductions using three dierent solvers (Glucose, Lingeling and PicoSAT) and the result are compared to that of MSieve, an open-source integer factorisationprogram. As seen in the comparison boolean satisability solvers cannot be used as a replacement of an integer factorisation solver. Additionally comparisons between the dierent reductions and boolean satisability solvers show that the combination of solver and reduction greatly aects performance. The implication is that further explorations of the integer factorisation problem with boolean satisability solvers can greatly benet from either avoiding a inadequate solver and reduction pair or from attempting to identify the outliers that signify a inadequate coupling.