Comparative Analysis of Traditional Time Series, Machine Learning, Deep Learning and Hybrid Models for Profit Forecasting in Financial Markets

Aug 4, 2026

Upload by Ophelia

This thesis compares the forecasting accuracy of traditional time series, machine learning, deep learning, and hybrid models for daily banking profit data in financial markets, aggregated at three levels. Method comparison uses a novel corrected mean average scaled error (cMASE) that improves interpretability over MASE by changing the one-step naive baseline so the naive benchmark consistently scores cMASE=1. Results show SARIMA as the most consistent performer, while hybrid approaches refine forecasts via long lags and rolling statistics with reduced refitting costs.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages73

Reading Time3 h 4 min

AccessFree to download

Introduction

Instruments in banking in financial markets area

Models

Traditional Time Series Models

Machine Learning Models

Deep learning Model

Hybrid Model

Loss functions

Importance and hypothesis

Methodology

Dataset

Data preprocessing

Modeling

Hyperparameter tuning

Results

Profit series aggregated on an instrument level

Profit series aggregated on a portfolio level

Profit series aggregated on a trading desk level

Discussion

Conclusions

Citations

University of Tartu Faculty of Science and Technology Institute of Mathematics and Statistics William Vaask Comparative Analysis of Traditional Time Series, Machine Learning, Deep Learning and Hybrid Models for Profit Forecasting in Financial Markets Actuarial and Financial Engineering Master’s Thesis (30 ECTS) Supervisor: Prof. Toomas Raus AEGRIDADE LINEAARSETE, MASINÕPPE, SÜVAÕPPE JAHÜBRIID MUDELITE VÕRDLEV ANALÜÜS KASUMI PROGNOOSIMISEL FINANTSTURGUDEL Magistritöö William Vaask Lühikokkuvõte Töös võrreldi lineaarsete, masinõppe, sügavõppe ja hübriid mudelite aegridade prognoositäpsust panganduses finantsturgude valdkonnas, prognoosideskolmel erineval tasemel agregeeritud igapäevast kasumit. Erinevate meetoditevõrdlemiseks kasutati uut jõudlusmõõdikut-korrigeeritud keskmine skaleeritud absoluutviga (cMASE), mis parandab MASE tõlgendatavust, kasutades T ühe-sammu naiivset prognoosi T − 1 asemel, mille tulemusel on naiivse meetodi skoor alati cMASE = 1 . Vaatamata arvutusvõimsuse arengust tingitud võimalikele mudelitele, oli lineaarne aegreamudel SARIMA teistest mudelitest parem, näidates kõige järjepidevamaid tulemusi keskmise ristvalideerimise cMASE ja testimise cMASEvahel erinevate aegridadega. Parimates hübriidmudelites täiendasid gradiendi võimenduse meetodid SARIMAt, parandades prognoose ajaliselt pikkade nihete, liikuvate statistikute abil. Kui SARIMA mudelid vajasid uuesti treenimist peale igat prognoosi, siismasinõppe, süvaõppe ja hübriid mudelite mittelineaarsed osad prognoosisid kõige paremini, kui neid treeniti uuesti ainult keskmiselt iga kahe nädala tagant, mis vähendas oluliselt kogu arvutuskulu. CERCS teaduseriala: P160 Statistika, operatsioonianalüüs, programmeerimine, finants-ja kindlustusmatemaatika. Märksõnad: MASE, SARIMA, XGBoost, LightGBM, LSTM, Hübriid. COMPARATIVE ANALYSIS OF TRADITIONAL TIME SERIES, MACHINE LEARNING, DEEP LEARNING AND HYBRID MODELS FOR PROFIT FORECASTING IN FINANCIAL MARKETS Master thesis William Vaask Abstract This thesis compared the forecasting performance of traditional time series, machine learning, deep learning and hybrid models on daily banking profit data in financial markets area aggregated on three different levels. To evaluate different methods, this thesis used a novel performance metric - corrected mean average scaled error (cMASE), which improves interpretability of MASE by using T one-step naive forecasts instead of T − 1, which results in naive method always having a score of cMASE = 1 . Despite various models available due to advancements in computational power, traditional time series method SARIMA still outperformed other models, showing the most consistent results between average cross-validation cMASEand testing cMASE. For best hybrid models, gradient boosting complemented SARIMA by correcting forecasts using long lags and rolling statistics. While SARIMA models required refitting after every forecast, machine learning, deep learning and non-linear parts of hybrid models performed best when refit only on average once every two weeks, which reduced the overall computing cost significantly. CERCS research specialisation: P160 Statistics, operations research, programming, financial and actuarial mathematics. Key Words: MASE, SARIMA, XGBoost, LightGBM, LSTM, Hybrid. Contents Introduction 5 1 Instruments in banking in financial markets area 6 2 Models 9 2.1 Traditional Time Series Models ................. 10 2.2 Machine Learning Models .................... 11 2.2.1 LightGBM ......................... 14 2.2.2 XGBoost .......................... 15 2.3 Deep learning Model ....................... 17 2.3.1 Finding optimal weights and biases ........... 19 2.3.2 Backpropagation algorithm ............... 20 2.3.3 Recurrent neural network ................ 22 2.3.4 Long short term memory ................. 23 2.4 Hybrid Model ........................... 26 2.5 Loss functions ...........................

FAQs

What problem does the thesis address?

The thesis addresses how to forecast profits in banking within financial markets using traditional, machine learning, deep learning, and hybrid modeling approaches.

How does the thesis evaluate forecasting methods?

It uses a corrected mean average scaled error (cMASE), designed to improve MASE interpretability by using T one-step naive forecasts so the naive method always achieves cMASE=1.

Which model family performs best overall and how do hybrid models help?

SARIMA delivers the most consistent results across cross-validation and testing. For top hybrid models, gradient boosting improves SARIMA forecasts by correcting them using long lags and rolling statistics.

Aug 4, 2026

Upload by Ophelia