---
title: Smooth Threshold GARCH: Quasi Maximum Likelihood Estimation and Empirical Evidence
url: https://www.ml-quant.com/papers/arxiv/2610.40047/
site: ML-Quant (https://www.ml-quant.com)
updated: 2026-10-30
license: Summaries CC BY 4.0; links go to the original sources
index: https://www.ml-quant.com/llms.txt
identifier: arXiv:2610.40047
source_url: https://arxiv.org/abs/2610.40047
featured: 2026-10-30
citations: unknown
topic: Derivatives & Volatility
---


# Smooth Threshold GARCH: Quasi Maximum Likelihood Estimation and Empirical Evidence

A GARCH variant with an estimated threshold captures volatility responding asymmetrically to positive versus negative returns, allowing positive returns to have no volatility impact.

- Source: https://arxiv.org/abs/2610.40047
- Identifier: arXiv:2610.40047
- Released: 2026-10-30
- First featured: Quant Letter No. 137 (2026-10-30): https://www.ml-quant.com/issues/2026-10-30/
- Citations (Semantic Scholar): not tracked
- Published in: not yet
- Topic: Derivatives & Volatility
- Authors: Jonas Fiedler Frederiksen, Frederik Vilandt

## Abstract (arXiv, CC0)

It is well established that the volatility of daily stock returns responds asymmetrically to news, as described by the news impact curve (NIC), which relates volatility to the previous return. Empirical studies suggest that (i) volatility reacts more strongly to negative returns than to positive returns of the same magnitude, (ii) the point of asymmetry, where the NIC attains its minimum, is positive, and (iii) returns above this point often do not increase volatility at all. The popular GJR and A-GARCH models each rule out one of these features. We study the Smooth Threshold GARCH (STGARCH) model, which nests both and allows for (i)-(iii) by combining an estimated threshold, the point of asymmetry, with separate slopes on either side of it. Although the likelihood is not twice continuously differentiable in the threshold, we show that the quasi maximum likelihood estimator is consistent and converges at the usual $\sqrt{n}$-rate, whether or not (iii) holds. If (iii) does not hold, the estimator is asymptotically normal, while if it holds, the slope above the threshold is zero, a value on the boundary of the parameter space, and the limit is non-Gaussian. In that case, the $t$-statistic for (iii) converges to the positive part of a standard normal variable, and the $t$-test of the GJR model remains valid. These results allow us to test the GJR and A-GARCH models, as well as (i)-(iii), within the STGARCH using standard normal critical values. For daily returns on five major stock indices, we reject the GJR model for all five and the A-GARCH model for three, suggesting that the STGARCH improves upon both. Volatility is lowest after returns 0.3-0.6 percentage points above their conditional mean, in line with (i) and (ii). Finally, the hypothesis in (iii) is rejected for two of the five indices.

## Related

- [Unified GARCH-Recurrent Neural Network in Financial Volatility Forecasting](https://www.ml-quant.com/papers/arxiv/2504.09380/): The paper proposes a new GARCH-GRU model for financial volatility forecasting, showing better computational efficiency and forecasting accuracy than other models.
- [Is the difference between deep hedging and delta hedging a statistical arbitrage?](https://www.ml-quant.com/papers/arxiv/2407.14736/): The research compares deep hedging and delta hedging in a GARCH-based market model, suggesting that the difference between the two can be a statistical arbitrage if the risk measure doesn't adequately consider negative outcomes.
- [Combining Deep Learning and GARCH Models for Financial Volatility and Risk Forecasting](https://www.ml-quant.com/papers/arxiv/2310.01063/): The research introduces a hybrid method for predicting the volatility and risk of financial tools by merging GARCH time series models with deep learning neural networks, finding that while this approach improves volatility predictions, it doesn't necessarily enhance Value-at-Risk and Expected Shortfall forecasts.
- [Improving Realised Volatility Forecast for Emerging Markets](https://www.ml-quant.com/papers/ssrn/4584573/): A study comparing four models for forecasting volatility in emerging markets found the HAR model best for long-term volatility and realised GARCH models for volatility clustering and persistence.
- [Candlestick Words Improve Intraday Volatility Forecasts Beyond HAR and GARCH](https://www.ml-quant.com/papers/ssrn/7621579/): Encoding intraday bars as descriptive words adds 1.7 to 3.7 points of out-of-sample R² to standard volatility models after spikes but shows no edge on daily bars.
- [Structured SDEs for Realized GARCH: Continuous-Time Volatility Modelling and Component-Level Analysis](https://www.ml-quant.com/papers/arxiv/2610.17732/): A continuous-time volatility model incorporating realized measures forecasts one-day-ahead integrated variance, showing strong out-of-sample performance relative to GARCH and related benchmarks across 17 equity indices.