X International conference�“Information Technology and Implementation” (IT&I-2023)�Kyiv, Ukraine

1

Research on the Use of Machine Learning Methods for Forecasting Time Series when Making Management Decisions in IT Projects Under Martial Law

​

Viktor Morozov, Vladyslav Deineha and Andrii Khlevnyi,

Taras Shevchenko National University of Kyiv, 24, Bohdan Gavrilishin Str., Kyiv, 04116, Ukraine

Dedicated to the tenth anniversary of the Faculty of Information Technology

Introduction

• Time series forecasting is an important area of research in data analysis, scientific modeling, and machine learning. This topic has become especially relevant in the context of the modern information society under martial law, where data accumulation is occurring at a tremendous speed, and the ability to identify patterns in the dynamics of time series for the needs of IT project management is becoming a key issue for many industries.

​

• In a rapidly changing business environment and uncertainty associated with various factors, such as economic and political events, time series forecasting is becoming an indispensable tool for making strategic management decisions. Thus, research in this area not only enriches our theoretical knowledge, but also has practical value, contributing to the development of more accurate and adaptive forecasting methods.

Use of time series for business forecasting

• Time series forecasting in business is one of the elements of planning and effective management. Predicting future trends helps companies adapt to changes in demand, market changes, and other influencing factors. In the field of inventory management, time series forecasts allow you to optimize inventory levels, avoiding excessive costs or shortages of goods.

​

• The use of time series forecasting methods in business is becoming especially important in the context of modern technologies and availability of big data. Machine learning methods and statistical models enable more accurate analysis and forecasting of time series dynamics, providing flexible and adaptive management of IT projects in a changing business environment. Thus, time series forecasting is becoming one of the key tools for sustainable and successful business development.

Analysis of recent research and publications

• Research and publications on the topic of time series forecasting were analyzed.
• The main focus was on considering the following methods: ARIMA, Exponential smoothing, Hybrid method.
• The main focus of the study was on the practical application of forecasting methods, as well as evaluation of the effectiveness of the methods under conditions of different data structures.The methods successfully demonstrated their effectiveness when applied to real data.
• These results emphasize the applicability of the methods under real-world conditions and their ability to successfully forecast time series in various practical applications.

Exponential Smoothing and ARIMA

• Exponential Smoothing is one of the basic approaches to time series forecasting. This method is based on the weighted consideration of previous values of the time series using exponential coefficients. The basic idea is that each observation of the time series is assigned a certain weight, which decreases exponentially with the increase of time backward.

​

• ARIMA (Autoregressive Integrated Moving Average) is a time series forecasting method that combines three main components: autoregression (AR), integration (I), and moving average (MA). This method was developed to model complex time series structures, including trends and seasonality. Its flexibility and adaptability make it a powerful tool for forecasting a variety of time series.

MAPE

• Before describing the hybrid method, it is important to mention the accuracy metric that will be used as an accuracy metric for forecasting methods.
• MAPE (Mean Absolute Percentage Error) is a metric for assessing the accuracy of forecasts, especially in the context of time series. This metric measures the average absolute percentage error between actual and forecast values. A low MAPE value indicates a high forecast accuracy.

​

MAPE = (1 / n) * ∑[( |Y(t) - Ŷ(t)| ) / |Y(t)| ] * 100,

where n - sample size; Y(t) - represents the actual value of the time series at time t; Ŷ(t) - represents the forecast value at time t.

Hybrid method (simple approach)

Hybrid method (improved approach)

Algorithm for building a time series forecast

• The following algorithm for building a time series forecast will be used.The main process is divided into several steps:
• Data splitting: The original time series is split into training and test samples.
• Model training: On the training sample, forecasts are made for each of the three methods. Each method forecasts time series values based on training data.
• Comparison and selection of the main model: The forecasts of each method are compared with the corresponding values of the test set to evaluate their accuracy on the test data. The main model is selected based on the MAPE metric, preferring the model with the “best” MAPE score.
• Forecast on all data: The selected main method forecasts time series values over the entire data range, which is the input time series.

​

Practical demonstration

• For a practical demonstration of building a time series forecast, specific data from the Kaggle website were selected, which are daily time series for the last 5 years for various indicators of website traffic with educational materials on statistical forecasting.
• For forecasting purposes, the indicator with the daily number of downloaded pages is selected. From this time series, the first 60 values are selected, on the basis of which the forecast for the next 7 values will be built, thereby representing a 7-day forecast.
• In this case, Python, various libraries for working with data and statistical analysis, as well as the Jupyter Notebook development environment were used to analyze and build a time series forecast.

Data preparation and visualization

• After loading the necessary libraries and time series data, the first step is to visualize the time series.
• Next, the time series is divided into two sets: training set and test set. The training set is used to train the model, while the test set is used to evaluate the accuracy of the forecasts.

Input time series

Building exponential smoothing and ARIMA models

• In the next step, we build models using exponential smoothing and ARIMA methods based on the training sample. The forecast values obtained using the constructed models will be further compared with the test data, allowing us to evaluate the accuracy of each method and select the best forecasting method for this time series.

Exponential smoothing model on the training set

MAPEs of two models

• After generating forecasts using the models, we proceed to calculate their MAPEs. MAPE values are calculated for each model, providing a quantitative assessment of the difference between forecasts and actual values.

​

• Corresponding MAPE values:
• MAPE_Exponential is 2.965060577251055
• MAPE_ARIMA is 6.107361536607362

Building a hybrid model

• Moving on to the hybrid model building, we implement two approaches: the first simple approach with equal weights, and the second improved approach with MAPE-dependent weights.

​

• After building the hybrid models, the corresponding MAPE values are calculated:
• MAPE_HybridSimple is 4.329272723465154
• MAPE_HybridMAPE is 3.7134184220187154

​

• After calculating the MAPE values for the two hybrid approaches, it can be clearly seen that the improved version, which takes into account the weights depending on the MAPE of each method, showed more accurate and efficient forecasts compared to the simple approach with equal weights. The results confirm that taking into account the performance of each method in forming the weights leads to improved accuracy of the hybrid model.

Building the final forecast

• Considering the MAPE metric results for each method, exponential smoothing showed the lowest error value. Based on this, it was selected as the main forecasting method. Then, based on this method, the final forecast is constructed over the entire time series, providing optimal accuracy and reliability in forecasting.

Initial time series and forecast values

Conclusions

• The paper examines the relevance of time series forecasting in various fields including finance, economics, healthcare and logistics. Methods such as exponential smoothing and ARIMA model are investigated with reference to research and publications. Forecasting methods including exponential smoothing, ARIMA model with automatic parameter selection and their hybrid combination were used in the practical part.
• A hybrid model with an improved approach using MAPE-dependent weights for each method was built, which showed a better result compared to the simple approach with equal weights in the practical part.
• Experiments on training and test data confirmed the superiority of the exponential smoothing method in minimizing prediction error (MAPE). This result allows to conclude that this method is applicable for a specific time series in this case.
• The study provides practical guidelines for the application of time series forecasting methods in real-world scenarios, emphasizing the importance of choosing the best method for managerial decisions and effective resource management.

Thank you for your attention!