Anyone selling an AI price chart should put the random walk's 1.294% error beside the neural network's 1.285%. Those are the mean absolute percentage errors for next week's FTSE 100 level in this paper. The gap hardly sells the chart.
Bhatti, Umer and Iqbal turn each return forecast into a price by multiplying last week's close by exp(predicted return/100). The arithmetic explains the visual result. Every model's return RMSE falls within 0.998 to 1.044 of the random walk's 1.717% on the FTSE 100, so every predicted price path hugs the actual path. At 144.3 points, the random walk's FTSE 100 price RMSE is within 0.23 points of anything that improves on it. For the FTSE 250, its 427.4 points is the lowest price RMSE among all eight models. The paper gets this exactly right: "a chart of predicted against actual prices cannot distinguish skill from persistence."
Eight models share a walk-forward test
Weekly FTSE 100 and FTSE 250 bars from Investing.com cover January 2015 to October 2026. The training sample contains 364 weekly returns through 2021. The test starts in January 2022 and runs for 251 weeks, taking in the 2022 rate shock and April 2025 tariff sell-off. All models are re-estimated weekly using an expanding window.
Two benchmarks, the random walk and historical mean, sit alongside an ARIMA. BIC selects its order on the training sample; that order then stays fixed. The authors also fit AR(1)-GARCH(1,1) and AR(1)-GJR-GARCH(1,1), each with Student-t errors. GJR allows negative shocks to increase variance more than positive shocks do. The remaining models are an OLS predictive regression, a 300-tree random forest and a 16-8 MLP averaged across five seeds.
All three learning models use the same 12 predictors. These comprise lagged returns, the absolute return, rolling volatility, 4/12/26-week momentum, the high-low range, volume change and the other index's return. Hyperparameters were set in advance.
Risk supplies the paper's positive result. In the full-sample GJR fits, the positive-shock coefficient reaches its zero bound. The asymmetry term is 0.246 for the FTSE 100 and 0.244 for the FTSE 250: falls account for almost all of volatility's response. Both halves of the story will be familiar to readers of Welch and Goyal or Hansen and Lunde, whom the paper cites. The authors claim three contributions. They offer a 2022 to 2026 UK sample, put machine learning and GARCH through a shared walk-forward return test (scoring GARCH alone on volatility), and demonstrate the illusion of accuracy discussed above.
Can 251 weeks exclude return skill?
Against the random walk, every RMSE ratio falls between 0.998 and 1.044. Only the neural network produces positive out-of-sample R² against the historical mean in both indices. Its gains are 0.13% and 0.14%, neither significant. The authors describe the edge as "caution, not insight," a fair judgement. L2 penalisation and early stopping drew its forecasts toward zero.
The random forest does worse than tell us nothing. Its FTSE 100 RMSE reaches 1.793% versus 1.717%, with DM = 2.79 (p < 0.01) and out-of-sample R² of −9.28%. It gets direction right in 44.2% of weeks; always predicting up scores 55.8%. The sole significant directional result belongs to ARIMA. Its FTSE 100 hit rate is 58.6% (Pesaran-Timmermann p = 0.017), just 2.8 points above always-up.
The abstract is careful: "none beats the random walk significantly." The conclusion goes further. Asked whether AI beats simple benchmarks, it says the answer "is no." The forest and predictive regression earn that judgement, with DM 2.79 and 2.36, both significantly worse. The neural network's verdict depends on a failure to reject.
With 251 weekly observations, Diebold-Mariano and Clark-West have little power to detect an R² gain of 0.13%. The scope of AI is also narrow here: one forest and one small MLP using 12 price-and-volume features. The authors acknowledge the price, range and volume inputs as a limitation. They leave LSTMs and time-series foundation models for future work. These data fail to find skill. They cannot exclude it.
GJR-GARCH's FTSE 250 Value-at-Risk clears Kupiec at p = 0.137
Volatility is the paper's stronger result. Using QLIKE loss against a Parkinson range proxy, which estimates volatility from the price range, GJR beats RiskMetrics EWMA on the FTSE 250: 0.414 versus 0.492 (DM −2.93, p < 0.01). The FTSE 100 comparison is 0.420 versus 0.445, with an insignificant DM −1.08. On the FTSE 250, GJR forecasts have a 24.5% squared correlation with realised range volatility, compared with EWMA's 3.4%. The FTSE 100 figures are 13.2% against 2.0%.
The abstract claims "correctly calibrated Value-at-Risk for both indices". Its FTSE 250 result gives less comfort than that wording suggests. At 95%, GJR records 18 FTSE 250 breaches against 12.6 expected, a rate of 7.2%. Kupiec returns p = 0.137. It passes. The FTSE 100 pass is clean, with 12 breaches against 12.6 expected at 95% (p = 0.873).
Symmetric GARCH, by comparison, has 20 FTSE 250 breaches (8.0%) and fails at p = 0.046. Two breaches mark the distance between calibrated and rejected. At 99%, FTSE 250 GJR VaR records 5 breaches against 2.5 expected (p = 0.164).
Twice, the paper turns non-rejection into a verdict: "no" for AI returns and "correctly calibrated" for Value-at-Risk. The same 251 weeks, with the same low power, support both judgements.
Kupiec counts breaches and ignores their timing. The paper describes only this unconditional test; we did not find an independence or conditional-coverage test. The authors' interval chart gives a reason to want one. Its 12 FTSE 250 misses cluster around abrupt moves: Ukraine in February 2022, the gilt-crisis rebound, March 2023 and April 2025. A desk sizing margin cares about exactly that clustering.
After 0.10% costs, FTSE 100 timing trails buy-and-hold
Buy-and-hold in the FTSE 100 earned 7.93% annually with a Sharpe of 0.65. All timing strategies fell behind after the 0.10% switching cost. ARIMA did best at 5.89% (Sharpe 0.58). The random forest lost 4.20% a year.
For the FTSE 250, buy-and-hold returned 0.66% and suffered a −27.78% drawdown. ARIMA returned 1.97% (Sharpe 0.18, drawdown −11.22%), while the predictive regression earned 1.36% (drawdown −20.08%). The authors attribute this to "risk reduction in a falling market": the strategies spent 44 to 45% of weeks in cash. I agree with that reading.
The authors concede that these simulations use price indices, pay zero on cash and omit dividends. The omissions work in opposite directions. Zero cash returns understate the strategies that stayed out for nearly half the weeks. Excluding dividends understates always-invested buy-and-hold. We found no turnover figures.
We could not rerun the study.
Our data do not include FTSE 100 and FTSE 250 OHLCV histories. The design requires both, since each index's return helps forecast the other. Its test period also extends to October 2026, beyond our history.
The useful change for a trader is in evaluation: test return forecasts formally against a random walk and disregard price-level MAPE. The risk result supports a narrower claim than advertised. It offers a significant QLIKE advantage on one index and a FTSE 250 VaR pass with 18 breaches against 12.6 expected. A conditional-coverage test showing that those 18 FTSE 250 breaches are not clustered in the 2022 sell-off would change my view.