📈 economics

The Illusion of Accuracy: Machine-Learning and GARCH Forecasts of UK Equity Index Prices

This study finds that while advanced machine learning and GARCH models accurately predict UK equity index prices, they fail to outperform simple benchmarks in return prediction, suggesting their true value lies in risk measurement rather than forecasting returns.

Original authors: Uzair Bhatti, Sadaat Umer, Awais Iqbal

Published 2026-10-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Uzair Bhatti, Sadaat Umer, Awais Iqbal

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of finance, there is a persistent hope that a clever algorithm can look at the past and see the future. This hope rests on the idea that stock prices, while seemingly chaotic, might hide patterns that a computer can learn to exploit. The central concept here is the "random walk," a simple benchmark that suggests the best guess for tomorrow's price is simply today's price, because all known information is already baked into the current number. If a sophisticated computer model cannot consistently predict price changes better than this simple guess, then the market is behaving efficient, and the promise of easy profits from prediction is an illusion. This question matters deeply to anyone with savings or a pension, because if artificial intelligence can truly forecast the market, it could reshape how wealth is managed; if it cannot, then relying on these tools might lead investors astray.

A recent study set out to test this promise directly, focusing on the United Kingdom's two major stock market indices. The researchers gathered weekly data from January 2015 through October 2026, a period that included significant economic turbulence such as interest rate shocks and tariff sell-offs. They pitted eight different forecasting methods against each other. On one side were the simple benchmarks: a model that just repeats the last week's price, and another that uses the average return of all past weeks. On the other side were the heavyweights of modern finance: traditional statistical models designed to track volatility, and three distinct types of artificial intelligence, including a random forest, which builds many decision trees to find patterns, and a neural network, which mimics the way a brain processes information. The goal was not just to see which model looked best on a chart, but to determine if any of them could actually predict the direction of the market better than chance, and if those predictions could turn into real money after paying for trading fees.

The results revealed a striking disconnect between how accurate the models looked and how useful they actually were. When the researchers compared the predicted price levels to the actual prices, every single model appeared highly accurate. The average error was tiny, ranging from 1.3 to 1.6 percent. On a graph, the predicted lines sat almost perfectly on top of the real lines, creating a visual impression of mastery. However, this accuracy was an illusion created by the nature of stock prices, which tend to stay close to where they were the week before. Because the prices were so persistent, even the simplest model, which just assumed nothing would change, achieved nearly the same level of accuracy as the complex artificial intelligence systems. When the researchers stripped away the price level and looked strictly at the weekly changes in value, the picture changed completely. None of the eight models, including the advanced machine learning systems, managed to beat the simple random walk in a statistically meaningful way. In fact, for the larger index, the random forest and a standard predictive regression performed significantly worse than simply assuming the market would stay flat.

The study also examined whether these models could help an investor make more money. The researchers simulated a trading strategy where an investor would buy the index if the model predicted a rise and hold cash if it predicted a fall, accounting for the cost of every trade. In the rising market of the larger index, the "buy and hold" strategy, which involves doing nothing and simply keeping the investment, outperformed every single forecast-based strategy after costs were deducted. The most sophisticated machine learning model actually lost money compared to just holding the stock. For the smaller index, some models did slightly better than buying and holding, but this was largely because they managed to stay out of the market during a sharp decline, a form of risk reduction rather than a sign of predictive skill. The researchers concluded that the apparent success of these models in predicting price levels was misleading; they were not forecasting the future, they were merely tracking the past.

While the models failed to predict where the market was going, they succeeded remarkably well at predicting how much it might swing. The study found that a specific type of statistical model, known as an asymmetric GARCH model, was far superior at forecasting volatility, which is the size of the price swings, rather than the direction. This model correctly anticipated that bad news would cause larger jolts in the market than good news, a phenomenon known as the leverage effect. It provided the most accurate estimates of risk and produced reliable "Value-at-Risk" numbers, which tell investors how much they could potentially lose in a bad week. This distinction is crucial: the tools were excellent at measuring the danger of the market but useless at predicting its destination.

The authors argue that this finding exposes a common trap in financial reporting. Many studies and commercial tools present charts where predicted prices closely match actual prices as proof of their power. This paper demonstrates that such charts are deceptive because they do not account for the fact that prices naturally stick together. The real test of a forecasting tool is whether it can predict the change in price better than a simple guess, and in this rigorous test, the artificial intelligence models failed. The value of these quantitative tools lies not in telling investors when to buy or sell to beat the market, but in helping them understand and manage the risk of their investments. For the curious investor, the lesson is clear: the ability to draw a line that follows the market is not the same as the ability to see where the market is going next.

Technical Summary: The Illusion of Accuracy in Machine-Learning and GARCH Forecasts of UK Equity Index Prices

Problem Statement
The paper addresses a critical gap in financial forecasting literature: the lack of rigorous, out-of-sample comparisons between modern machine-learning (ML) models and traditional econometric benchmarks within the UK equity market. While AI is increasingly marketed as a superior tool for predicting stock prices, much of the supporting evidence relies on visual comparisons of predicted versus actual price levels or low percentage errors. The authors argue that because stock prices are highly persistent, a simple forecast (such as repeating the previous week's price) will naturally appear accurate on a price chart, masking a lack of genuine predictive skill. The study seeks to determine whether AI models can genuinely outperform simple benchmarks (the random walk and historical mean) and econometric models (ARIMA, GARCH) in predicting UK index prices, returns, and volatility, and whether any such statistical gains translate into economic value after transaction costs.

Methodology
The study employs a rigorous, expanding-window, out-of-sample design using weekly data for the FTSE 100 and FTSE 250 indices from January 2015 to October 2026. The sample is divided into a training period (2015–2021) and a test period (2022–2026), covering significant market events including the 2022 interest-rate shock and the 2025 tariff sell-off.

Eight distinct models are evaluated:

Benchmarks: Random Walk (RW) and Historical Mean (HM).
Econometric Models: ARIMA, AR–GARCH, and AR–GJR-GARCH (incorporating asymmetric leverage effects and Student-t errors).
Machine Learning/Statistical Learning Models: Predictive Regression (OLS), Random Forest (RF), and a Feed-forward Neural Network (MLP).

All learning models utilize 12 predictors, including lagged returns, volatility measures, momentum, and volume. Forecasts are generated weekly, re-estimating parameters to prevent look-ahead bias. Evaluation is multi-faceted:

Price Accuracy: Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE).
Return Predictability: Out-of-sample $R^2$ , Diebold-Mariano (DM) tests, Clark-West tests (for nested models), and Pesaran-Timmermann direction tests.
Bias Analysis: Mincer-Zarnowitz regressions testing for unbiasedness ( $\alpha=0, \beta=1$ ).
Volatility and Risk: Comparison against RiskMetrics (EWMA) and 12-week rolling variance using MSE and QLIKE loss functions; Value-at-Risk (VaR) backtesting via Kupiec tests.
Economic Value: Simulation of trading strategies (long when predicted return > 0, cash otherwise) with 0.10% transaction costs, compared against a buy-and-hold benchmark using Sharpe ratios and maximum drawdown.

Key Results
The study yields three primary findings that challenge the prevailing narrative of AI superiority in price prediction:

The "Illusion of Accuracy" in Price Levels: All models, including complex ML architectures, produced price forecasts with Mean Absolute Percentage Errors (MAPE) of 1.3–1.6%. However, this accuracy is driven almost entirely by the persistence of price levels. The Random Walk benchmark performed as well as or better than all complex models. Notably, the Random Forest and Predictive Regression models were significantly less accurate than the Random Walk for the FTSE 100. No model significantly beat the Random Walk in terms of return predictability or direction (hit rate), and Mincer-Zarnowitz regressions confirmed that predicted returns explain almost none of the variation in actual returns ( $R^2 < 1.22\%$ ).
Superiority of GARCH in Volatility Forecasting: In stark contrast to return prediction, the asymmetric GJR-GARCH model significantly outperformed historical benchmarks in forecasting volatility. It provided the most accurate volatility forecasts for both indices and produced correctly calibrated Value-at-Risk (VaR) estimates at 95% and 99% confidence levels. This confirms that while the direction of prices is unpredictable, the magnitude of price moves (risk) is forecastable, particularly when accounting for the leverage effect (asymmetry).
Lack of Economic Value: No forecast-based trading strategy outperformed a simple buy-and-hold approach for the FTSE 100 after accounting for transaction costs. For the FTSE 250, some strategies showed slightly higher returns, but this was attributed to risk reduction (staying out of the market during declines) rather than predictive skill, as these models failed to demonstrate significant return predictability in statistical tests.

Key Contributions
The paper makes three specific contributions to the literature:

UK Market Evidence: It provides rare, recent out-of-sample evidence for the UK market, extending a literature dominated by US data.
Comprehensive Comparative Design: It is one of the few studies to directly compare AI and GARCH-family models side-by-side across four distinct criteria: price accuracy, return predictability, volatility forecasting, and economic value within a single, robust experimental design.
Methodological Clarification: It demonstrates that visual comparisons of predicted versus actual price levels are insufficient for evaluating forecasting models. The paper establishes that the "illusion of accuracy" arises from price persistence and argues that the random walk must be the minimum benchmark for any claim of predictive skill.

Significance and Claims
The paper concludes that the value of quantitative models in UK equity markets lies in measuring risk rather than predicting returns. The authors assert that the apparent success of AI models in price charts is often a statistical artifact of market persistence rather than evidence of exploitable inefficiency.

The study suggests that investors and regulators should be skeptical of AI tools that present price-level charts as proof of accuracy. Instead, the standard for evaluation should be statistical significance against a random walk and economic viability after costs. The findings support the view that while AI and ML may not currently offer a "crystal ball" for market direction, GARCH-family models remain essential tools for risk management, position sizing, and setting stop-loss levels. The paper modestly claims that its results serve as a practical test for evaluating future claims regarding AI forecasting tools, particularly those utilizing large language models or foundation models.

Technical Summary: The Illusion of Accuracy in Machine-Learning and GARCH Forecasts of UK Equity Index Prices

More like this