ILLUSTRATION: SHUTTERSTOCK
NEPAL’S STOCK EXCHANGE has been through three technological ages. Traders shouted and scribbled until screen-based dealing arrived in 2007. The NEPSE Online Trading System followed in 2018. Full automation came in January 2021. Saurav Karki, an MBA at King’s College and an official at Naasa Securities, has now taken daily index returns from July 1995 to February 2025 and asked a basic question in his paper in the NRB Economic Review: does the market price things sensibly and does it do so all the time? By mid-January 2025 the exchange listed about 267 firms and turned over roughly NPR 5.21bn a day, mostly in small trades by retail investors.
The answer is that it does, sometimes, and then it stops. Earlier studies of NEPSE usually relied on monthly figures or short samples and concluded that the market was inefficient in the weak sense: past prices could predict future ones. Karki uses daily data, cuts the record into two-year slices and runs five-year rolling windows over it. Persistent inefficiency from 1999 to 2019 gives way to tentative signs of efficiency from 2021 to 2025, alongside financial digitisation and regulatory reform. That fits the adaptive market hypothesis put forward by Lo, which says efficiency is not a fixed property but something investors learn and lose as conditions change.
The linear tests agree with that story. Over the full sample, returns show clear serial dependence and mean reversion. But the subsamples move around. There was a strange quiet patch in 1997 and 1998, when political instability and collapsing trading volumes seem to have wiped out the patterns that arbitrageurs normally exploit. Once normal trading resumed, inefficiency returned and stayed for two decades. Then, in the most recent period, the old predictability fades again. Rolling windows pick out the same shape: a brief surge in efficiency in the late 1990s, then a long stretch of the opposite, then a recent softening.
Non-linear tests are less forgiving. Volatility clustering—the tendency of wild days to follow wild days—shows up in every period. Higher-order dependence, the sort of complex pattern that comes from herd behaviour and feedback trading, does not budge. In the most recent data, the simpler volatility effects ease a little, but the deeper complexity stays. Information may be flowing better than before; the way investors react to it has not changed as much.
The GARCH models, which estimate how shocks feed into volatility, put persistence at about 0.90. That is high. Shocks do not fade quickly in Kathmandu. Volatility spikes appear during the Asian financial crisis and the Maoist insurgency, the global financial crisis, the earthquake and the 2015 Indian trade blockade, the pandemic and the bull market that followed it, and the monetary tightening of 2022 and 2023. Calm stretches in 2010-2013 and 2017-2019 are just as visible.
One result stands out. Unlike developed markets, NEPSE does not punish bad news more than good. Volatility rises by roughly the same amount whichever way prices move. The usual explanation for the asymmetry elsewhere—from leverage to loss aversion to mechanical selling—does not apply in a market with little margin trading and mostly long-only retail investors. Fear of missing out and panic selling both lift variance, and they lift it equally.
The Markov-switching model finds two states. In the calm one, daily volatility is around half a percentage point and returns drift sideways. In the turbulent one, volatility is about twelve times higher and returns are mildly positive, because turbulence in Nepal tends to come with speculative rallies rather than pure declines. Calm periods last about fifteen trading days. Turbulent ones last about nine. The model classifies the whole 2020-2021 bull market as turbulent, which makes sense: it tracks uncertainty, instead of direction.
Efficiency differs between the two states. Turbulent periods show more predictability and more complex dependence. Calm periods come closer to a random walk. Momentum and mean-reversion strategies that work in one state can fail in the other. The quiet efficiency of 1997-1999 came with such slim trading that it was hardly a model for anything. The recent improvement comes with automated settlement, tighter oversight, more DEMAT accounts and financial literacy programmes. Linear predictability weakens while volatility clustering persists. Information processing improves at one speed, behaviour at another.
Comparisons abroad fit parts of the picture. Pakistan shows similar volatility persistence. India, China and Pakistan have all seen efficiency shifts after financial reform. Finland shows that small size does not guarantee inefficiency; structural reform helps, with a lag. Latin America shows calendar anomalies appearing and disappearing with conditions. Nepal shares the pattern. Efficiency moves with competition and institutions, and not with the calendar alone.
Three decades of daily trading end with signs of improvement and old habits intact. That combination, rather than a clean arrival at developed-market standards, is the finding worth carrying forward. ■
