A Data Analysis and Market Price Prediction of Ethiopian Commodity Market with Machine Learning Algorithms
No Thumbnail Available
Date
2018-03
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
AAU
Abstract
The current Ethiopian market is conducted in a traditional manner and market drivers are still
not used for prediction of future market price. Although, large amount of market data have been
gathered throughout years by both governmental and non-governmental organizations, yet little
have been done to analyze the data for future market price prediction. Moreover, the analysis
methods were often manual creating inefficiency in time and quality of market prediction.
Analyzing valuable data will show us what the future holds and accelerate the development goals
of the country in the sector. The study examines features of current Ethiopian market attributes
to find out most valuable features for predicting market price. Eighteen technical indicators are
taken and tested for their individual ability of prediction and redundancy. From the feature
selection of commodity marke, we have found that features like Stochastic %K, Stochastic %D,
Close gain/loss, High, close price, Opening Price, Low, RSI, Ton and Moving Average
Convergence/ divergence (MACD) founded to be in the top ten of individual performance
evaluation. Moreover features namely Stochastic %K, Relative Strength Index (RSI), Bollinger
Bands-Upper, Highest-High, close gain/loss, Simple Moving Average (SMA), Closing price,
MACD-Fast, Exponential Moving Average (EMA), MACD-Slow and Low founded to be less
redundant. The study also compares four machine learning models for their prediction ability of
Ethiopian commodity market price. The outcomes of feature selection were used to compare the
models. Two experiments were conducted; the first was comparison of the models with 10 fold
cross validation using feature of high individual predictive ability and less redundancy. The
second one was a comparison of models with separate train and test data using features of high
individual predictive ability and less redundancy. From the models (Support Vector Machine
(SVM), Artificial Neural Network (ANN), K-Nearest Neighbor (K-NN) and Ensemble Learning)
the performance of ANN and Ensemble Learning algorithms are shown to be accurate than SVM
and K-NN. The average MAE rate of the ANN model was 2.8084. Ensemble Learning and SVM
follow with average MAE rate of 4.9362 and 8.1178 respectively. The other model was least
performer with the MAE rate above 45.3381.
Description
Keywords
Feature selection, Technical Indicators Price prediction, Machine Learning Algorithms