Financial time series prediction under Covid-19 pandemic crisis with Long Short-Term Memory (LSTM) network

The COVID-19 pandemic triggered unprecedented disruptions in global supply and demand, heightened uncertainty, and materially impacted financial markets worldwide. Traditional portfolio theory relies on historical data and assumes known future returns, but forecasting errors during crises can severely impact investment strategies. Predicting asset prices is therefore crucial for portfolio allocation and risk management, particularly under crisis conditions. This study aims to design and implement an LSTM neural network to forecast prices across multiple financial asset classes—global and regional stock indices, U.S. government bond indices, and a comprehensive set of S&P GSCI commodity indices—over an extended period (1998-01-02 to 2020-09-16) that includes several market crises, with special attention to performance during the COVID-19 period. The purpose is to assess whether LSTM can capture temporal dependencies in financial time series and provide accurate forecasts to inform diversification and investment decisions.

Prior work shows neural networks, including LSTM, often outperform linear models in predicting financial series. Studies on Chinese and other markets (Cao et al., 2005; Chen & Li, 2006) found neural networks can improve forecasting quality versus linear and STAR models. LSTM has been shown to effectively model nonlinear dependencies and provide high-accuracy predictions for stocks, exchange rates, and commodities (Chen et al., 2015; Fischer & Krauss, 2018; Rundo et al., 2019). Hybrid and enhanced architectures (ARIMA-CNN-LSTM, LSTM-CNN with attention) further reduce errors for assets like carbon futures and gold (Ji et al., 2019; He et al., 2019). Comparative studies generally report LSTM outperforming ARIMA and other machine learning models (Siami-Namini et al., 2018; Nikou et al., 2019), while some find GRU or tailored RNNs competitive in specific tasks (Sako et al., 2022; Zaheer et al., 2023). During COVID-19, multivariate CNN-LSTM models have shown strong performance for stock indices (Widiputra et al., 2021). Overall, literature supports LSTM’s suitability for financial time-series forecasting, motivating its application across varied asset classes and stress periods.

Data: Daily closing prices from 1998-01-02 to 2020-09-16 sourced from Datastream. Assets include equity indices (MSCI World, MSCI US, MSCI Europe, MSCI Pacific, MSCI EAFE, MSCI Emerging Markets), U.S. government bond indices (S&P US 5–10Y, US Benchmark 10Y), and 34 S&P GSCI (Total Return) commodity indices across energy, industrial metals, precious metals, agriculture (softs and grains), and livestock. Preprocessing: Min–max normalization scaling each series to [0,1]. Each index is split 80% for training and 20% for testing. The training span covers multiple crises (e.g., 1998, 2001, 2008, 2011–2012), enabling the model to learn behavior across calm and turbulent phases; the test span includes the COVID-19 period. Model architecture and training: A univariate LSTM per index with a single input layer (one input variable), one hidden layer comprising four LSTM blocks, and a single-output layer. Sigmoid activation is used. Weights and biases initialized randomly; training for 100 epochs. Optimization employs the ADAM algorithm, minimizing standard regression losses. Evaluation: Performance assessed using root mean squared error (RMSE) and mean absolute error (MAE) on both train and test sets, with an emphasis on test-set performance for generalization. Comparative benchmarks: ARIMA-type models per index, selected via minimum AIC and estimated by maximum likelihood, evaluated with the same metrics over identical train/test splits.

• Overall performance on global indices: Using normalized data, average train RMSE and MAE were 19.72 and 13.84; average test RMSE and MAE were 18.19 and 13.60, indicating strong predictive performance.
• Equity indices: Most regional equity indices were predicted accurately (e.g., averages for non-US regions RMSE Test ≈ 17.46, MAE Test ≈ 12.73). The MSCI USA exhibited the highest test errors (RMSE Test 139.22; MAE Test 108.21), attributed to the sharp, speculative rise during COVID-19; the model captured trend direction but with reduced precision.
• U.S. government bonds: Very low errors, reflecting lower price variability. US Benchmark 10Y (RMSE Test 1.32; MAE Test 0.84) and S&P US 5–10Y (RMSE Test 0.37; MAE Test 0.30); bond average (RMSE Test 0.85; MAE Test 0.57).
• Energy sector commodities: Generally strong performance, with average RMSE Test 23.29 and MAE Test 19.79. Some series showed larger errors, notably Natural Gas (RMSE Test 80.82; MAE Test 80.82) and Unleaded Gasoline (RMSE Test 49.55; MAE Test 34.90); Biofuel showed the smallest errors (RMSE Test 1.88; MAE Test 1.58).
• Industrial metals: Average RMSE Test 11.29; MAE Test 8.48. For example, Copper (RMSE Test 42.69; MAE Test 31.97) was harder to predict, while Aluminum (RMSE Test 0.63; MAE Test 0.46) and Zinc (RMSE Test 1.89; MAE Test 1.44) had low errors.
• Precious metals: Average RMSE Test 11.50; MAE Test 7.73. Gold (RMSE Test 6.93; MAE Test 4.59) was forecast accurately despite safe-haven surges; Silver and Platinum also showed low errors.
• Agriculture: Softs exhibited the smallest errors (average Softs RMSE Test 2.44; MAE Test 2.06). Grains average RMSE Test 11.49; MAE Test 8.56. Livestock average RMSE Test 23.61; MAE Test 16.51.
• During COVID-19 tests, the LSTM captured decline in many commodity prices (with cocoa an exception), and the strong rise in gold and U.S. equities trends, demonstrating robustness under stress.
• Benchmark comparison: Across nearly all assets, LSTM outperformed ARIMA-type models on both train and test sets (lower RMSE/MAE). The notable exception was Natural Gas, where ARIMA achieved slightly better test performance (e.g., ARIMA RMSE Test 80.55 vs. LSTM 80.82).

The study set out to determine whether an LSTM architecture could accurately forecast prices across diverse financial asset classes, particularly during crisis periods like COVID-19. Results show the LSTM effectively models temporal dependencies and generates predictions close to observed prices across equities, U.S. bonds, and commodities. Lower errors for bonds and many commodities indicate strong generalization in less volatile or structurally consistent series; higher errors for MSCI USA reflect extreme, rapid upswings during COVID-19, where direction was captured but magnitude was harder to match. Compared with ARIMA, LSTM’s consistent error reductions confirm the advantage of nonlinear sequence modeling for financial time series. Practically, accurate multi-asset forecasts support portfolio diversification and hedging decisions, enabling investors to identify assets to include in optimal strategies during turbulent markets. The demonstrated resilience during COVID-19 underscores the model’s relevance for risk management under systemic shocks.

An LSTM-based forecasting framework was designed to predict daily prices for multiple financial asset classes—global and regional equity indices, U.S. government bond indices, and a broad set of commodity indices—over 1998–2020, including the COVID-19 period. The model captured temporal interdependencies and produced accurate forecasts across assets, with particularly low errors for bonds and many commodity sub-indices. Compared to ARIMA-type models, the LSTM delivered lower RMSE and MAE for nearly all assets, evidencing superior predictive capability. These results suggest LSTM forecasts can inform optimal diversification and investment strategies, even during crisis periods. Future work includes developing hybrid architectures (e.g., CNN–RNN) for greater accuracy and longer horizons, extending to digital asset markets, and modeling volatility by combining recurrent networks with GARCH-type approaches.

The study employs a univariate LSTM with a fixed architecture (single hidden layer with four LSTM blocks, sigmoid activation) and uniform preprocessing across diverse assets; alternative or deeper architectures, exogenous features, and hyperparameter optimization could further improve accuracy. The forecast horizon focuses on one-step daily predictions and a specific 80/20 split; other horizons and rolling evaluations might better characterize robustness. The dataset, while broad, excludes other asset classes (e.g., cryptocurrencies) and macro/alternative data that may enhance predictability. Natural gas results indicate certain series may require different models. Future research avenues include hybrid CNN–LSTM architectures, prediction of digital asset prices, and volatility forecasting via RNN–GARCH hybrids.