EFFICIENCY ANALYSIS OF FINANCIAL TIME SERIES FORECASTING MODELS UNDER MARKET TURBULENCE CONDITIONS

Abstract

This paper presents a comparative analysis of financial time series forecasting models' effectiveness under market turbulence conditions. The study focuses on evaluating the adaptability of statistical ARIMA model and recurrent LSTM neural network across different prediction horizons during periods of high market volatility. Daily OHLC data from five major technology companies (Google, Apple, Amazon, Meta, Oracle) for the period 2020-2025 was analyzed, with particular emphasis on the turbulent April-June 2025 period. Three model architectures were implemented: ARIMA(2,1,0), LSTM Bidirectional Autoencoder (100 units), and simple LSTM (20 units). Testing was conducted on 5, 15, and 30-day forecasting horizons using MAPE, RMSE, and MAE metrics. Additionally, residual analysis through autocorrelation function examination was applied to validate model quality. Results demonstrate that ARIMA excels in short-term forecasts (5 days) with MAPE ≤ 0.06, but its effectiveness diminishes on medium-term horizons due to inability to adapt to market turbulence. Simple LSTM (20 units) achieved optimal balance between accuracy and stability, outperforming ARIMA by 30.75% on medium and long-term forecasts. Complex LSTM Autoencoder proved least effective due to overfitting on market noise. The scientific novelty lies in comprehensive analysis of model adaptability to extreme market turbulence using residual analysis as additional validation method. It was proven that simpler LSTM architectures outperform complex ones under high volatility conditions. The practical significance includes optimization of algorithmic trading strategies and risk management systems during market instability periods, particularly valuable for financial institutions and investment funds.