Explore publications.

The purpose of this study was to investigate the efficacy of hybrid forecasting models that integrate the classical Autoregressive integrated moving average framework, the support vector machines, the artificial neural networks, and random forest for predicting S&P 500 index returns. Utilizing daily share price data from 2015 to 2023, the methodology used a two-stage decomposition process where ARIMA captured linear dependencies, while the ML models projected nonlinear patterns within the residuals. The Out-of-sample analysis revealed that all the hybrid models significantly outperformed their standalone ARIMA benchmark across key accuracy metrics, such as mean absolute error, root mean squared error, and directional accuracy. The ARIMA-RF hybrid achieved the highest predictive accuracy and statistical significance, though it entailed a greater computational cost. The findings confirmed that the hybrid forecasting  capabilities to mitigate linear autocorrelation and ARCH effects, offering a robust approach for financial time series forecasting.

This study investigates the application of Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) models for predicting Bitcoin prices, with the objective of enhancing investment approaches. Given Bitcoin's significant volatility and inherent risk, accurate price forecasting is vital for making strategic decisions. The research evaluates model performance by analyzing Bitcoin price data spanning May 25, 2020, to May 25, 2024, employing quantitative methodologies like experimental design and predictive analysis. Results indicate the LSTM model achieves superior predictive accuracy over the RNN model, demonstrated by reduced Mean Absolute Percentage Error (MAPE) and Root Mean Squared Error (RMSE) figures. Additionally, the research delves into how these models affect the Dollar Cost Averaging (DCA) investment strategy, offering perspectives on the advantages and drawbacks of using machine learning to optimize cryptocurrency investments. 

This study examines whether Large Language Models (LLMs) can support momentum-based breakout strategies for retail investors by analysing technical charts and validating statistical patterns. We analysed 3,621 price charts (2023–2025) to assess LLMs’ ability to identify trading setups and deviations from random walk behaviour. 13,624 tests were run on 1,204 trades using various statistical tests. While 70% of patterns conformed to random walk behaviour in the short term (7–15 days), momentum signals rose to 26.42% over longer periods (50–250 days). Complex prompting produced both higher random walk classification (72.41% vs. 69.29%) and better returns than single-shot prompting. Incorporating Gamma Exposure (GEX) improved caution but reduced return rates. Discrepancies of up to 20% between JavaScript and Python implementations underscored LLMs’ current limitations in statistical precision. Overall, LLMs may democratise technical analysis but still require human oversight, structured prompting, and validation through traditional methods.

The purpose of this study is to determine which indicators are more capable of showing more accurate sell and buy signals on the LQ45 index by using the oscillator indicator Moving Average Divergent Convergent (MACD), Bollinger Band, and Relative Strength Index (RSI). The results in this study indicate that the sell signal can be captured well by the Bollinger band and MACD indicators, but it cannot be captured properly by the RSI, the volume can be small or heading and are in the side ways, while the MACD plays a too slow role in capturing the signal buy compared to Bollinger bands and RSI. The use of a single indicator will never show a buy and sell signal that is really accurate, this is based on the results of research that shows the difference in timeliness in Bollinger, RSI and also MACD so that the combination of several types of indicators will be better compared to using single indicators. Although in statistics there are no significant differences, there are only differences in increment and improvement in the placement of existing values, but in this case, the order of this value is crucial for traders because it requires very high accuracy to determine the right decision in daily transactions.

A logistic regression model is has also become a popular model because of its ability to predict, classify and draw relationships between a dichotomous dependent variable and dependent variables. On the other hand, the R programming language has become a popular language for building and implementing predictive analytics models. In this paper, we apply a logistic regression model in the R environment in order to examine whether daily trading volume at the Botswana Stock Exchange influence daily stock market movement. Specifically, we use a logistic regression model to find the relationship between daily stock movement and the trading volumes experienced in the recent five previous trading days. Our results show that only the trading volume for the third previous day influence current stock market index movement. Overall, trading volumes of the past five days were found not have an impact on today’s stock market movement. The results can be used as a basis for building a predictive model that utilizes trading as a predictor of stock market movement.

Univariate Time Series Models [TSM] use only a Panel of historical data to produce forecasts. The tacit belief in using TSM is that the past information portends the future of the longitudinal data-stream. This is likely in certain cases such as strictly Ergodic Panel segments of sufficient size in the overall Panel. A question of interest is: Is the success of TSM in these contexts generalizable? The test of this question used a Litmus-Test design to examine the performance profile of TSM for a longitudinal time series the last point of which is a Turning Point [TP]. Specifically, the inference measure will use the Relative Absolute Error [RAE] of the TSM tested over three forecasting horizons. In this testing, five TSM configurations were employed; the TPs are identified using a fixed screening filter applied to randomly selected firm Panels actively traded on the S&P500 from 2005 through 2013.  There is no evidence that any of the five TSM outperformed the RW model which is incidentally the TP. The impact of these results is that one cannot assume that the effectiveness of TSM generalizes to all domains—in particular—forecasting after TPs that seems to be a Domain Lacuna where the effectiveness of TSM will be compromised.

The purpose of this study is to explore the importance and ranking of technical analysis variables in Turkish banking sector. Random Forest method is used for determining importance scores of inputs for eight banks in Borsa Istanbul. Then two predictive models utilizing Random Forest (RF) and Artificial Neural Networks (ANN) are built for predicting BIST-100 index and bank closing prices. Results of the models are compared by three metrics namely Mean Absolute Error (MAE), Mean Square Error (MSE), Median Absolute Error (MedAE). Findings show that moving average (MAV-100) is the most important variable for both BIST -100 index and bank closing prices. Therefore, investors should follow this technical indicator with respect to Turkish banks. In addition ANN shows better performance for all metrics.

The aim of this study to examine the performance of Support Vector Regression (SVR) which is a novel regression method based on Support Vector Machines (SVM) approach in predicting the Istanbul Stock Exchange (ISE) National 100 Index daily returns. For bechmarking, results given by SVR were compared to those given by classical Linear Regression (LR). Dataset contains 6 technical indicators which were selected as model inputs for 2005-2011 period. Grid search and cross valiadation is used for finding optimal model parameters and evaluating the models. Comparisons were made based on Root Mean Square (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Theil Inequality Coefficient (TIC) and Mean Mixed Error (MME) metrics. Results indicate that SVR outperforms the LR for all metrics.