Rohit Kumar

Interestingly, I met at least four young people who have put their careers on hold to be part of what one of them called “this fight for democracy.”

Interestingly, I met at least four young people who have put their careers on hold to be part of what one of them called “this fight for democracy.”

Beneath all their bluster at being always right, lies a fear of being challenged. As young Indians refuse to blindly follow the status quo, they must decide whether to cling to old beliefs or reevaluate them.

Beneath all their bluster at being 'always right,' lies a fear of being challenged. As young Indians refuse to blindly follow the status quo, they must decide whether to cling to old beliefs or reevaluate them.

Listen to this article: The ‘WhatsApp uncles’ are worried. Like the October weather, the social and political climate around them is beginning to change and they don’t quite know what to make of it. The youth protests burgeoning across India over the...

There is something to be said about the quiet, hopeful wisdom of those who have witnessed and participated in many battles for democracy.

Listen to this article: New Delhi: On October 23, human rights activists, student unions, farmers and other citizens came together to protest under the banner of ‘People Against SIR’ in the heart of Delhi. Their attempt to march to the Election...

Abstract Telomeres protect chromosome ends and support genomic stability through interactions among telomerase, shelterin proteins, DNA-repair factors, and chromatin regulators.

There is something to be said about the quiet, hopeful wisdom of those who have witnessed and participated in many battles for democracy.

It tests the forecasting accuracy of Random Forest and XGBoost for the next-day log-return forecasting for four equity markets of BRICS countries: Brazil (IBOVESPA), China (SHCOMP), India (NIFTY 50), and South Africa (JSE Top 40) on a sample of 4,087 observations per country between January 2010 and September 2025. Russia has been omitted due to restrictions on data for sanctions which prohibit the construction of an equivalent series over the estimation window. Four methodological changes are made in the design from the original submission. The first one is that the next-day log return is used as the target variable instead of the next-day closing price; this eliminates the mechanical R2 inflation from LAG_1 in the features. Second, four baseline forecasts, including naïve persistence, AR(1), auto-selected ARIMA, and GARCH(1,1) for the volatility arm (in Appendix C) are estimated with the machine learning models. Third, an expanding window walk-forward scheme with a 5-year training window and a 1-year validation window and a 1-year step is used in lieu of a fixed train/test split. Fourth, the significance of pairwise forecast comparisons is assessed with the Diebold–Mariano statistic and the Harvey–Leybourne–Newbold small-sample correction. Three different feature-space specifications are used: one for market-technicals-only (24 features), another with macro variables from both the home country and the United States (92–101 features), and a third with only the economically-motivated macro variables (26–27 features). The main empirical result is that starting from the returns target, AR(1) is statistically superior to XGBoost at p < 0.001 in all three feature-space specifications and all four markets. XGBoost R2 is negative in all cases: − 0.1378 for India, − 0.0571 for China, − 0.0684 for Brazil, − 0.0578 for South Africa, and, as the feature space increases, worsens in all four cases for the full macro-financial specification: − 0.2730, − 0.1872, − 0.2521, and − 0.1133. Random Forest also fails to beat AR(1) at 5% significance in the technical specification, but under the full macro specification for South Africa attains R2 of 0.0010 and a Diebold–Mariano p value of 0.9176 against AR(1)—the one country-specification pair where an ML model is not statistically dominated by AR(1). A similar hyperparameter-tuned LSTM (with grid search over hidden units and sequence length) underperforms even more, yielding R2 values from − 2.28 (South Africa) to − 7.05 (India), a factor of 10–100 smaller than those of XGBoost, and with Diebold–Mariano tests showing a lack of equal expected loss with AR(1) at p < 0.0001 in all four markets (Appendix D). There is no model consistently able to achieve more than 50% directional accuracy. In fact, a Bai–Perron structural-break analysis of the Indian return series suggests that a shock occurred when the NIFTY 50 price index hit a break in April 2020, with the mean level of the post-break regime being at least 80 per cent higher than the pre-break regime, which leads to the conclusion that in-sample-trained models can’t generalise. Volatility of the rolling return is the top feature channel across all markets, while the US Federal Reserve policy rate ranks as the most important external factor for Brazil (14.8 per cent of the SHAP contribution) and the South African Reserve Bank policy rate for South Africa (9.5 per cent of the SHAP contribution). The findings suggests the weak-form market efficiency hypothesis and the recent literature which cautions against specious assertions of machine learning’s superiority in equity return prediction.