MSc.SS Theses and Dissertations (2022)

Permanent URI for this collection

http://hdl.handle.net/11071/13167

Browse

Now showing 1 - 10 of 10

A Comparative study of Hybrid Neural Network and ARIMA Models with application to forecasting intra-day child-line calls in Kenya
(Strathmore University, 2022) Wang’ombe, Grace Wairimu
Background: For successful staffing and recruiting of call centre professionals, precise forecasting of the number of calls arriving at the centre is crucial. These projections are needed for various periods, both short and long-term. Benchmark time series models such as ARIMA and Holt-Winters used in forecasting call centre data are outperformed in long term forecasts, especially when the data is not stationary. Advanced models such as the ANNs can pick up on the random peaks or outlying periods better than the benchmark time–series models. The hybrid methodology combines the strengths of the benchmark time–series and advanced models, thus improving overall forecasts. Objective: The study’s primary goal was to assess the superiority of a Hybrid ARIMAANN model over its constituent models in forecasting Childline call centre data in Kenya. Methods: The ARIMA, ANN and hybrid ARIMA-ANN models were used in the call centre data forecasting. The cross-validation technique was used to create forecasting accuracy metrics which are then compared. In ARIMA, the Box-Jenkins methodology is used to fit the model whereas the neural network element of the hybrid model and the ANN were modelled using the feed-forward Neural Network Autoregressive(NNAR) structure. Results: The Seasonal ARIMA - ANN model outperformed the ARIMA model in short term forecasts and the ANN model in long term forecasts. The Diebold-Mariano test indicated a significant difference between the hybrid and ANN forecasts, whereas the difference between the hybrid and ARIMA forecasts was not significant. Conclusion: The Hybrid model was able to adapt both of its constituent models’ advantages to better its performance. These results are helpful as call centres can be able to use one model which is robust enough to create accurate forecasts rather than the benchmark models.
A Systematic comparison of performance of Ridge, Lasso, Elastic net and Relaxed Elastic net when fitting high dimensional data for sales prediction
(Strathmore University, 2022) Muoki, Monica Mueni
Forecasting or prediction is one of the most crucial aspects of planning for many companies. Data-driven decisions can only be as accurate as the prediction they are based on. Some of the decisions include production planning, inventory management, and various resource allocation. Sales information is really multi-dimensional, and as a result not easy to analyse. Our motivation is to reduce the high dimension of this information, select optimal contributing variables with the aim of making accurate and reliable sales predictions. The purpose of this study is to compare the performance of four restricted regressions. This involves looking at Ridge, Lasso, Relaxed net and Elastic net regressions and assessing their performance in prediction when dealing with high dimensional data. The proposed method will involve comparison of the four mentioned regularized techniques, citing their restrictions and evaluating their prediction model performance. We will also involve data simulation to test the different models. The simulations are done under different scenarios to present the reality of a market setting. Afterwards, we will select the best model and use it to fit our real sales dataset provided by one of the leading ECMCs in Kenya. On this basis, elastic net offered best predictions based. The evaluating metrics for this models are Root Mean Squared Error (RMSE), Mean Absolute Error (MAE) and R-Squared (R2). However, the desired model based on R2 kept shifting under different scenarios to Lasso, Ridge and Elastic net. The results indicated that the regularized approaches especially elastic net are capable of dealing with non-linearity and fluctuating dynamics in manufacturing industry while predicting electrical cable sales accurately.
Assessing predictive performance of supervised machine learning algorithms: an alternative model for diamond pricing
(Strathmore University, 2022) Kigo, Samuel Njoroge
The world’s hardest mineral is a diamond, which is 58 times harder than any other mineral, and its beauty as a jewel has long been appreciated. The diamond is popular due to its optical property as well as other causes such as its durability, custom, fashion, and strong marketing by diamond producers. Diamond demand, on the other hand, is not directly related to such inherent characteristics, but rather to their perceived value as rare and expensive objects. Forecasting diamond pricing is challenging due to non-linearity in important features such as carat, cut, clarity table, and depth. Given this, we conducted a comparative analysis and implementation of multiple supervised machine learning models in predicting diamond price in both classification and regression approaches. We evaluated eight different supervised algorithms in our work, including Multiple Linear Regression, Linear Discriminant Analysis, eXtreme Gradient Boosting, Random Forest, k-Nearest Neighbors, Support Vector Machines, Boosted Regression and Classification Trees, and Multi-Layer Perceptron, and showcased the best suitable model given selected evaluation metrics. The analysis in this work is based on data preprocessing, exploratory data analysis, training the aforementioned models, assessing their accuracy, and interpreting their results. Based on the performance metrics values and analysis, it was discovered that eXtreme Gradient Boosting was the most optimal algorithm in both classification and regression, with a R2 score of 97.45% and an Accuracy value of 74.28%. As a result, the eXtreme Gradient Boosting method was recommended for forecasting the price of a diamond specimen.
Comparison of neural networks and tree-based ensemble methods in detecting correlates of breast cancer survival
(Strathmore University, 2022) Katam, Ruth Jepchirchir
Breast cancer is common among women impacting about 2.1 million women each year, and causing a big number of cancer-related deaths. Most times doctors have a struggle in diagnosing the stage to determine accurately and needed medication. Therefore, accurate detection of correlates of breast cancer survival is paramount. This study sought to compare the performance of Neural Networks and Tree-based Ensemble methods to predict breast cancer survival, elucidating on factors causing breast cancer based on clinical data for timely intervention. The accuracy score, recall score, precision score, Area under Receiver- Operating Characteristic Curve, and F1 score were used to evaluate the performance of each model in discerning between breast cancer survivors and non-survivors. XGboost and LSTM exhibited an outstanding performance in the classification of Breast cancer patients. However, XGboost was the most optimal model. The results depicted that age at diagnosis, pam50+ claudin low subtype her2, 3 gene classifier subtype high, profile,radiotherapy,Nottingham prognostic index,type of breast surgery breast conserving, type of breast surgery mastectomy, mutation count, lymph nodes examined positive, tumor stage, tumor size, 3 gene classifier subtype low profile, pre inferred menopausal state and Post inferred menopausal state. among others were the most important correlates of survival from breast cancer.
Forecasting the term structure of interest rates in Kenya using Bayesian models post 2007-2008 financial crisis
(Strathmore University, 2022) Bosire, Luycer Nyanchama
Despite the growing significant advances in the modelling of the term structure of interest rates after the great recession of 2008, little attention has been paid to the problem of forecasting the term structure which has proven to be an important rate in several products and instruments offered by financial institutions. This dissertation makes use of a Dynamic Nelson-Siegel model with a Time-Varying Vector Auto- Regressive component to fit a model and forecast the h-step ahead expected yield. The model makes use of four parameters representing a decay factor, level, slope and curvature latent factors estimated with high efficiency. We propose to use our DNS-TV-VAR model to estimate our factors and demonstrate the model consistency to a range of stylized yield curve initial data. We apply the model in forecasting a term structure for short and long horizons and conclude that the forecasts appear more accurate for long horizons.
Improving performance of hurdle models using Rare-Event Weighted Logistic Regression: application to maternal mortality data
(Strathmore University, 2022) Okello, Sharon Awuor
Hurdle models, which are commonly used alongside zero-inflated models to analyze dispersed zero-inflated count data, employ a logit link function to predict whether an observation takes a positive count or a zero count based on a set of covariates. However, the logit model tends to be biased toward the majority zero class in cases involving rare events, and may underestimate the positive counts when their proportion is significantly smaller than that of the zero counts. This research aimed to improve the performance of hurdle models by incorporating rare-event weighted logistic regression model. Poisson and Negative Binomial (NB) Hurdle Rare Event Weighted Logistic Regression (REWLR) model estimates were developed and fit on various simulation conditions and maternal mortality data for performance evaluation using Akaike Information Criterion (AIC) and Area Under Curve (AUC). The Negative Binomial Hurdle REWLR emerged to be the best performing among all the evaluated models due to the ability to handle dispersion and adjust for class imbalance. The research findings will provide reliable estimates of the maternal mortality ratio in Nairobi without the risk of over-fitting zero counts.
Machine learning based prediction of life expectancy
(Strathmore University, 2022) Lipesa, Brian Aholi
The social and financial systems of many nations throughout the world are significantly impacted by life expectancy (LE) models. Numerous studies have pointed out the crucial effects that life expectancy projections will have on societal issues and the administration of the global healthcare system. These approaches offer a variety of strategies to enhance society-related advanced care planning and healthcare. Over time, research has proven that the vast majority of the existing factors were insufficient to forecast the lifespan of the general population. An understanding of the chosen sampling population’s death rate served as the foundation for earlier models. Researchers have asserted that despite improvements in forecasting approaches and meticulous work in the past, there are still several elements that must be taken into account to determine life expectancy rates in addition to death rates. As a result, life expectancy research now includes a broader focus on issues related to education, health, the economy, and social welfare. In this study, the author developed a model for estimating life expectancy rates taking into consideration health, socioeconomic, and behavioral characteristics by using the eXtreme Gradient Boosting (XGBoost) algorithm to data from 193 UN member states. The effectiveness of the model’s prediction was compared to that of the Random Forest (RF) and Artificial Neural Network (ANN) regressors utilized in earlier research. XGBoost attained an MAE and an RMSE of 1.554 and 2.402, respectively. It outperformed the RF and ANN models that achieved MAE and RMSE values of 7.938 and 11.304, and 3.86 and 5.002, respectively. The overall results of this study support XGBoost as a reliable and efficient model for estimating life expectancy.
Modeling of count data with an informative time component in the presence of overdispersion
(Strathmore University, 2022) Owiti, Levi Alfred Orero
In real-world count data, several methods have been applied to handle the common problem of overdispersion. However, these methods have not comprehensively considered unique features that may exist in the data. This study sought to address robust statistical modelling of count response data that contains temporal features. The study proposed a Bayesian Negative Binomial model that will handle over dispersion while taking into account the temporal features of the data. Two count data models were compared and extended to incorporate an informative time component. To test the various models, this study conducted simulation studies under specified parameters to examine how the models behave under certain conditions. The data generation mechanism ensured the simulated data had seasonality as is with the real-world data on fire frequency, temperature, and rainfall. Further, the study examined the effect of the additional components on prediction intervals of the simulation studies for the different count models. The introduction of Bayesian techniques into the modeling was intended to create more accurate prediction intervals that take account of the prior distribution of the data. The Bayesian Negative Binomial model was better than the Negative Binomial model in terms of model bias. When validated on real data to confirm its effectiveness, the Bayesian model had better MASE and the prediction intervals enveloped the actual data in the testing dataset of fires in Kenya between the year 2000 and 2018.
Statistical learning for class imbalanced data: a case study of Malaria indicator survey data
(Strathmore University, 2022) Ongera,Maangi Daniel
Class imbalanced problems are predominant in real-life applications. In most cases, the minority class is the most important. Standard statistical learning algorithms tend to produce poor results for the minority class and very good results for the majority class. One of the widely used mechanism to address this problem is by re-sampling the training data. The objective of this study is to examine the performance of statistical learning algorithms by using different re-sampling approaches for handling class imbalance. Methods Two classical and ensemble statistical learning techniques were trained on an imbalanced Malaria Indicator Survey data set while handling the majority-minority problem through re-sampling. These included: Logistic regression, support vector machines, random forest, and extreme gradient boosting. The algorithms were trained without handling class imbalance first. Secondly, the algorithms were trained using six re-sampling procedures to handle class imbalance: random under-sampling, random over-sampling, Synthetic Minority Oversampling technique (SMOTE), Random Over Sampling Examples (ROSE) techniques and Adaptive Synthetic Sampling Approach (ADASYN). We further investigated whether combining randomly under-sampled and over-sampled data can result in improved performance. Eighty percent of the data was used for model training using 5 fold cross validation. Results All methods that were considered for handling class imbalance had strengths and weaknesses depending on the performance metric. For instance, random under-sampling (RU) resulted in models with higher sensitivity than random over-sampling (RO). To get a trade-off between sensitivity and specificity, these two methods can be combined (RURO). This approach resulted in 99.5% sensitivity, 88.1 % specificity, 89.6 % precision, 94.3 % F1 score and a 93.9 % accuracy on the test set using the Extreme Gradient Boosting machine.
Temporal-difference comparison of learning methods for stock market prediction
(Strathmore University, 2022) Maina, Stephen Gakuo
Background: a stock/securities exchange is considered to be among the primary indicators’ of a country’s economic strength and development. Stock market prices are volatile in nature and are affected by factors like inflation, economic growth, etc. Prices depend heavily on demand and supply dynamics. Stock market price determination using ANNs has gained a lot of traction lately due to the obvious advantages this would represent to traders. Most methods in use today have largely been based on the feed forward algorithms, however, evolutionary techniques remain largely unexplored for this process despite their obvious robustness. Method: Using data from the Nairobi Securities Exchange, and specifically the NSE20 share index, the project will seek to apply and compare traditional ANN techniques for stock market prediction against the relatively new evolution algorithms. The Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE) and a confusion matrix will be calculated for performance evaluation. Results: the empirical results showed that the proposed evolutionary techniques out performed classic artificial neural networks methods-feed forward backpropagation.

Browse

Browsing MSc.SS Theses and Dissertations (2022) by Title

Results Per Page

Sort Options