When the COVID-19 pandemic sent global markets into a tailspin in early 2020, investors worldwide scrambled to understand what was happening to their portfolios. Stock prices that had followed predictable patterns for years suddenly lurched in unexpected directions. For researchers studying how to forecast stock movements, the pandemic created a natural experiment: which prediction models could adapt to unprecedented market chaos, and which would fail?

A new comparative study analyzing over a decade of stock data from the Bombay Stock Exchange reveals that not all forecasting models are created equal. While traditional time series methods excel at short term predictions during stable periods, deep learning models, particularly multilayer perceptron networks, demonstrated superior accuracy and resilience across both calm and turbulent market conditions.

The research examined eight different forecasting approaches, from simple moving averages to sophisticated neural networks, testing each against two datasets: one ending before the pandemic, and another including the volatile COVID-19 period. The findings offer practical guidance for investors, traders, and financial institutions navigating an increasingly unpredictable market landscape.

The Challenge of Predicting Market Movements

Stock prices reflect a complex web of influences spanning macroeconomic forces, industry trends, and company specific developments. Inflation rates, unemployment figures, and GDP growth shape the broader environment. Corporate events, from mergers to leadership changes, can trigger sharp price movements. Even market psychology, including overconfidence and herd behavior, plays a measurable role.

Traditional forecasting relies on two primary approaches. Technical analysis examines historical price patterns and trading volumes to identify trends. Fundamental analysis evaluates economic indicators and company financials to estimate intrinsic value. Yet both approaches face a fundamental challenge: markets are not fully predictable, and the factors driving prices constantly shift.

The emergence of machine learning and deep learning has introduced new possibilities. These algorithms can identify complex, nonlinear patterns in vast datasets that escape human detection or traditional statistical methods. The question is not whether these advanced techniques work in theory, but which ones perform best under real world conditions, especially during market stress.

Testing Models Against Market Reality

The researchers analyzed monthly closing prices from the Bombay Stock Exchange between 2010 and 2021, deliberately choosing this market for its concentration of small scale companies known for price volatility. This volatility makes it an ideal testing ground for prediction models.

The study divided data into training and testing sets. Models learned patterns from January 2011 through December 2018, then made predictions for the period from January 2019 forward. By comparing predictions against actual prices, researchers could measure each model's accuracy using multiple metrics: root mean square error, mean absolute error, and mean absolute percentage error.

Three categories of models underwent testing. Time series models included simple moving average, autoregressive integrated moving average (ARIMA), and Holt-Winters methods. Machine learning approaches featured support vector regression, random forest regression, and extreme gradient boosting (XGBoost). Deep learning was represented by long short term memory networks and multilayer perceptron models.

Each approach brings different strengths. Time series models assume future prices depend primarily on past prices, making them straightforward but limited in capturing complex relationships. Machine learning models can detect nonlinear patterns and select the most relevant features from noisy data. Deep learning excels at processing sequential information and adapting to shifting market dynamics.

Clear Winners and Losers Emerge

The results revealed stark performance differences. Among all tested models, the multilayer perceptron achieved the lowest error rates: a root mean square error of just 0.6 before COVID-19 and 0.7 when pandemic data was included. Its mean absolute percentage error remained below 1% in both scenarios, demonstrating exceptional stability.

Time series models showed clear limitations for long term forecasting. ARIMA, simple moving average, and Holt-Winters all struggled when market patterns broke during the pandemic. Their error rates increased substantially when COVID-19 data entered the picture. ARIMA's mean absolute percentage error actually improved slightly, from 11.60% to 9.46%, but its root mean square error more than doubled. These models work well for next month or next quarter predictions during stable periods but falter when markets enter uncharted territory.

Machine learning models presented a mixed picture. Support vector regression performed poorly, with the highest errors across the board. Its mean absolute percentage error nearly doubled during COVID-19, jumping from 8.85% to 17.48%. This sensitivity to market disruption reveals a critical weakness for a model that should theoretically handle nonlinear relationships.

In contrast, random forest regression and XGBoost demonstrated impressive resilience. XGBoost showed the most stability among machine learning approaches, with minimal changes in error metrics between pre pandemic and pandemic periods. Its mean absolute percentage error barely budged, rising only from 1.53% to 1.60%. These ensemble methods, which combine predictions from multiple decision trees, proved adept at generalizing to new market conditions.

Long short term memory networks, a type of recurrent neural network designed for sequential data, maintained consistent performance across both time periods. Their root mean square error held steady at 1.58, though the mean absolute percentage error increased modestly from 18.23% to 20.02%. The architecture's ability to selectively remember relevant historical information while discarding noise helped it weather the market storm.

What COVID-19 Revealed About Model Reliability

The pandemic period offered unique insights into model robustness. Before COVID-19, markets exhibited relatively stable conditions with predictable patterns and consistent trading volumes. Traditional factors like earnings reports and economic indicators drove price movements in familiar ways.

The pandemic shattered these patterns. Extreme volatility became the norm, with unprecedented price swings occurring within days. Trading volumes fluctuated wildly as investors alternated between panic selling and aggressive buying. Market sentiment shifted in response to health updates, policy announcements, and vaccine news rather than traditional economic fundamentals.

Models relying on historical pattern continuity struggled. Support vector regression, which maps input data to separate hyperplanes to maximize margins between data points, could not adapt quickly enough to the erratic movements. Time series models that assumed stable mean and variance over time found those assumptions violated repeatedly.

The multilayer perceptron's superior performance stemmed from its architecture. With multiple layers of interconnected neurons, each processing and transforming information, the model could learn complex, hierarchical representations of market behavior. It identified subtle relationships between variables that remained relevant even when surface patterns changed dramatically.

Random forest and XGBoost also benefited from their ensemble nature. By averaging predictions from many decision trees, each trained on different subsets of data, these models achieved robust generalization. No single tree's errors could dominate the final prediction, providing a buffer against outliers and unusual market conditions.

Matching Models to Investment Strategies

The findings have direct implications for different types of investors and trading approaches. For day traders and others focused on short term movements, simple moving averages and machine learning models like random forest or XGBoost offer reliable guidance. These methods capture recent trends and adapt quickly to changing conditions, providing timely signals for buy or sell decisions.

Long term investors face different requirements. They need models that can identify sustained trends and remain accurate across months or years, even when markets encounter unexpected shocks. Here, deep learning models, particularly multilayer perceptron networks, demonstrate clear advantages. Their low error rates and stability across different market regimes make them valuable for strategic planning and portfolio allocation.

Institutional investors managing large portfolios might benefit most from a diversified modeling approach. Using simple moving averages for short term tactical adjustments, machine learning for medium term positioning, and deep learning for long term strategic decisions allows each model type to contribute where it performs best.

The research also highlights an important trade off between accuracy and interpretability. Time series models are transparent; anyone can understand how a moving average works or what an ARIMA forecast assumes. Deep learning models operate more like black boxes, making decisions through millions of weighted connections that defy simple explanation. This opacity creates challenges for regulatory compliance and risk management, even as it delivers superior predictions.

The Broader Context of Financial Forecasting

These findings align with a growing body of evidence that machine learning and deep learning can extract meaningful signals from financial data. Previous research has demonstrated that neural networks can identify patterns invisible to traditional methods, that ensemble approaches improve reliability, and that sophisticated feature selection enhances prediction quality.

The study extends this work by providing direct, head to head comparisons under both normal and stress conditions. Most comparative analyses test models only during stable periods or use simulated data. By incorporating the COVID-19 shock as a natural experiment, this research offers rare insight into real world performance when it matters most.

The choice of the Bombay Stock Exchange adds another dimension. Many forecasting studies focus on major US markets or other developed economies. Emerging markets often exhibit different characteristics: less liquidity, higher volatility, greater sensitivity to global shocks. That deep learning models maintained their advantage in this more challenging environment suggests the findings may generalize broadly.

However, limitations remain. The study examined monthly closing prices, which smooth out intraday volatility and may miss relevant information. Future research incorporating higher frequency data, additional market indicators, or alternative data sources like social media sentiment could refine these results further.

The pandemic's unique nature also warrants caution. While it tested model resilience to extreme disruption, other types of market stress such as gradual structural changes or policy shifts might favor different approaches. No model can guarantee perfect predictions, and past performance, even under stress, never ensures future results.

Looking Ahead

As financial markets grow more complex and interconnected, the tools for understanding them must evolve as well. The superiority of deep learning models in this study suggests that artificial intelligence will play an expanding role in investment decision making.

Yet the research also underscores that no single model dominates in all scenarios. Time series methods retain value for specific applications. Machine learning approaches offer a middle ground between simplicity and sophistication. The optimal strategy likely involves combining multiple models, each contributing its particular strengths.

For regulators and policymakers, these findings raise important questions. If many market participants adopt similar advanced models, could that create new forms of systemic risk? When algorithms trained on similar data make similar decisions, markets might experience coordinated movements that amplify volatility rather than smooth it.

For individual investors, the message is both encouraging and cautionary. Better forecasting tools can improve returns and reduce risk. But relying too heavily on any model, however sophisticated, carries danger. Markets remain fundamentally uncertain, shaped by countless factors that no algorithm fully captures.

The COVID-19 pandemic reminded everyone that unprecedented events do occur, disrupting even the most carefully constructed models and strategies. The models that weathered this storm best were those built to handle complexity and adapt to change. As markets continue evolving, those same qualities of flexibility and robustness will remain essential, whether in algorithms or in the investors who use them.

Credit & Disclaimer: This article is a popular science summary written to make peer-reviewed research accessible to a broad audience. All scientific facts, findings, and conclusions presented here are drawn directly and accurately from the original research paper. Readers are strongly encouraged to consult the full research article for complete data, methodologies, and scientific detail. The article can be accessed through https://doi.org/10.1007/s42979-025-03848-y