Contributing Author · Quantitative Research
Niraj Neupane
CA (ICAI) · Quantitative Researcher · Financial Economist · Forward Deployed Engineer
Niraj works on the measurement of extreme loss in equity markets — how much a position can lose on a bad day, how confident anyone is entitled to be in that number, and what happens to it when liquidity disappears. His research sits at the join between classical financial econometrics and machine learning, and it is written to be validated rather than admired.
Background
Niraj is a Chartered Accountant under the Institute of Chartered Accountants of India (ICAI), and works as a quantitative researcher and financial economist. His professional focus spans quantitative trading, machine learning and AI applications in financial markets, financial econometrics, and model validation. He also works as a Forward Deployed Engineer — the part of the job where a model stops being a notebook and has to survive contact with a production risk system.
The accountancy training matters more to the research than it might appear. Risk models in regulated institutions are not judged on elegance; they are judged on whether they can be documented, challenged, backtested and defended to a supervisor. That is the standard his papers are written against, which is why they report the results that do not clear a significance threshold alongside the ones that do.
His current work covers one-day-ahead Value-at-Risk forecasting for US equities, conditional Extreme Value Theory in the tail, liquidity and volatility regime features, and the backtesting machinery — Kupiec, Christoffersen, Basel traffic light — that regulators actually use to decide whether a model is acceptable.
Published research
Working papers, available in full on SSRN.
SSRN Working Paper · Abstract 7222958
ML-LiqVaR: A Liquidity- and Regime-Aware, Extreme-Value-Calibrated Gradient-Boosting Value-at-Risk Model for US Equities
A hybrid framework for one-day-ahead VaR forecasting that combines GARCH volatility modelling, conditional Extreme Value Theory in the tail, and gradient-boosted quantile regression, conditioned on volatility, momentum, VIX and liquidity features. Tested on 8,072 out-of-sample observations across eight S&P 500 constituents, 2018–2024, against four benchmark models.
SSRN Working Paper · Abstract 7170418 · July 2026
Backtesting Value-at-Risk Models Under SR 11-7: A Comparative Analysis of Kupiec, Christoffersen, and Basel Traffic-Light Tests Applied to S&P 500 Returns (2018–2024)
A walk-forward backtest of Historical Simulation, Parametric Normal and Student-t VaR on S&P 500 daily returns, recalibrated every day on a trailing 250-day window and evaluated against the three tests a US supervisor applies under the Federal Reserve's SR 11-7 model risk management guidance.
Areas of work
Tail risk measurement
Value-at-Risk and Expected Shortfall, conditional Extreme Value Theory, and the question of what a 99% quantile estimate is actually worth when it is fitted on a few hundred observations.
Financial econometrics
GARCH-family volatility models, quantile regression, regime identification, and the long-standing question of how much of a return distribution is genuinely forecastable.
Machine learning in markets
Gradient boosting for quantile targets, feature construction from liquidity and volatility data, and the discipline of testing an ML model against the econometric benchmark it claims to beat.
Model validation
Backtesting under SR 11-7, coverage and independence testing, Basel traffic-light classification, and building models that can be documented and challenged rather than only deployed.
On this contribution
Niraj writes for Financial Gurkha as an independent contributing author. His articles here summarise his own published research and are not sponsored, commissioned by, or written on behalf of any employer, client or institution. Views are his own and do not represent those of any organisation he is affiliated with.
Research summarised on this site is working-paper stage. Working papers on SSRN have not been peer reviewed, and conclusions may change between drafts. Where a result is not statistically significant, we say so in the article rather than in a footnote.
Our sourcing, disclosure and corrections policy is set out in the Editorial Standards.