Too much of a good thing: Algorithmic trading, price discovery, and the efficiency threshold in emerging equity markets

Whether algorithmic trading (AT) improves or harms stock market efficiency is a central question in market microstructure research. This paper asks whether the AT–price discovery relationship is monotonic or whether a participation threshold exists beyond which AT intensity impairs price discovery. Using a quadratic entity fixed-effects panel model across 615 firms and 36,245 firm-month observations on India’s National Stock Exchange (NSE) over 2014–2024, with regulatory-grade firm-level AT intensity from Securities and Exchange Board of India (SEBI) mandatory order-tagging data, we identify an efficiency threshold at 38.65% AT intensity (95% confidence interval: [36.42, 40.88]). Below this threshold AT intensity reduces price delay and improves price discovery; above it, price discovery deteriorates. The nonlinearity is formally confirmed by the bilateral Lind and Mehlum (2010) U-shape test with a minimum absolute t -statistic of 4.97, exceeding the 1.645 critical value. The threshold declines from 34.71% for large-cap firms to 38.85% for mid-cap firms, while the efficiency-degrading arm is not yet identifiable for small-cap firms. A bid–ask spread mechanism is confirmed at a higher threshold of 43.94%, creating a 5.3 percentage-point early-warning window for regulators. Unlike prior linear approaches, this threshold-based framework reveals that 72.1% of firm-month observations already exceed τ ∗ , implying the prevailing consensus substantially underestimates the efficiency cost of algorithmic participation. Oster (2019) selection bounds and a lagged AT specification yield a consistent threshold of 38.72%, confirming robustness to omitted-variable concerns and reverse causality.