Stop-Loss strategies are often used by investors to combat negative returns by predetermining thresholds at which they should exit trades. Though existing traditional Stop-Loss mechanisms such as Fixed Stop-Loss and Trailing Stop-Loss are empirically proven to have the ability to minimize risks associated with trades, they still face serious challenges when it comes to achieving a balance between risk and return. In this study, we develop a Deep Learning model that combines the concept of Stop-Loss with the capabilities offered by Deep Neural Networks. The architecture is composed of three components where trend detection and price prediction components provide inputs to the stop price prediction component which predicts the variation of stop price. The study focuses on short term trades of minute frequency which involves analysing massive chunks of data that fluctuates rapidly within extremely short time intervals. Despite the advancements that has taken place in the context of Big Data analytics, intraday financial time series analysis has not received much academic attention. The model utilizes convolutional layers to capture spatial features along with Long Short Term Memory networks to capture temporal dependencies in price sequences. The proposed solution gives outstanding results for diverse market conditions and it works specifically well for trades that are downtrending. We evaluate our model for a portfolio that consists of five stock symbols from NASDAQ that are adequately liquid and two cryptocurrencies which are highly circulated. The model delivers a stable outcome across all symbols indicating its ability to be generalized over a range of symbols and its tolerance to diverse market conditions. Results also indicate that Stop-Loss mechanisms possess the potential to work well even in speculative markets such as the cryptocurrency market. The ability to reduce losses without compromising on opportunities to realize profits in dynamic and unstable market conditions is an important property of our Stop-Loss solution.

Modern application servers run on multi-core hardware and these servers use multi-threading to achieve high performance. Performance of such servers is highly dependent on the number of processing threads in the worker thread pool. The optimal thread pool size (that would result in the best performance) depends on workload properties (both the application and the incoming arrival pattern). In this paper, we propose ADAPT-T, a lightweight, adaptive algorithm that can auto-adjust the thread pool size to optimize the latency. ADAPT-T exploits the concave upward property of system performance and optimizes the latency on-line by periodically monitoring the server's behaviour and then adjusts the thread pool size based on monitored statistics. It has a special locking mechanism to minimize the frequent fluctuations that may occur in latency while tuning the thread pool size. We perform extensive experiments under different workload patterns and show that ADAPT-T performs well under a wide range of workload scenarios compared to other policies.

This work presents a classifier capable of classifying websites accessed by a community of users based on characteristics of web traffic they generate. The classifier introduced here can be used for engineering web traffic. This classifier is capable of identifying a few different classes of websites based on intensity of bursts of web requests made to each website in both the long and the short run. Previous studies have proposed different classification algorithms in the area of general network traffic classification and web traffic mining. Work presented in this paper is based on Kohonen's self organizing map trained using web traffic traces gathered from a busy proxy server. By integrating this classifier with a rate controlling module of a proxy server, web traffic can be engineered very easily.

A neural network model called the hierarchical overlapped neural gas (HONG) network is introduced and its performance on several datasets is described. In order to obtain improved classification accuracy, the HONG network partitions the input space by projecting the input data onto several different second layer neural gas networks. This duplication enables the HONG network to generate multiple classifications for every sample presented in the form of confidence values, and these confidence values are combined to obtain the final classification. Excellent recognition rates for several benchmark datasets are presented.

The paper describes our investigation into the neural gas (NG) network algorithm and the hierarchical overlapped architecture (HONG) which we have built by retaining the essence of the original NG algorithm. By defining an implicit ranking scheme, the NG algorithm was made to run faster in its sequential implementation. Each HONG network generated multiple classifications for every sample data presented as confidence values. These confidence values were combined to obtain the final classification of the HONG architecture. Three HONG networks based on three different feature sets with global and structural features were also trained to obtain better classification on conflicting handwritten data. An excellent recognition rate for the NIST SD3 database was consequently obtained.

This paper describes our investigation into the neural gas (NG) network and the hierarchical overlapped architecture which allowed us to obtain an excellent recognition rate for the NIST SD3 database. By defining an implicit ranking scheme, we made the NG algorithm runs faster in its sequential implementation. The hierarchical overlapped architecture allowed us to obtain multiple classifications for each sample data. Since a multiple classifier system is a powerful tool for difficult pattern recognition problems, we developed three classifiers based on three different feature extraction methods, with global and structural features.