Steve Yang

GARCH(1,1) outperformed flexible gate models under QLIKE because it separates level, persistence, and innovation share. We introduce PGARCH — a predictor-driven GARCH decomposition — and benchmark Linear PGARCH and XGB-g-PGARCH against the seven models from Part 4.

An introduction to the statistical tools for comparing forecast accuracy — the Diebold-Mariano test, the role of the loss function, and the Mincer-Zarnowitz regression as a calibration diagnostic.

How does parameter estimation error affect the Sharpe ratio of a mean-reversion strategy?

We keep RMSE for benchmark comparability, then refit STES and XGBSTES under QLIKE to see how the training loss changes out-of-sample ranking and calibration.

How does a correlated bias in the fair-value estimate affect the Sharpe ratio of a mean-reversion strategy?

What happens to the performance of a mean-reversion strategy when the fair value is estimated with a constant bias?

When centering (or standardizing) features leaves the intercept unchanged, and when it does not.

We augment STES and XGBoost-STES volatility models with rolling signature features computed from SPY return paths. After careful feature selection, regularization tuning, and multi-channel augmentation, we conclude that the results are negative but instructive.

We extend STES by replacing its linear component with XGBoost, then dig into why RMSE can underperform and propose a more stable end-to-end fitting algorithm.

Continuing from the previous post, we work through three examples of approximating path functionals using linear functionals of truncated signatures in increasingly less practical way to illustrate…

Previously I provided some motivations for the use of signatures.

I am starting a series on the signature method now that I have some free time on my hands over the 2025 holiday season.

In our previous post, we introduced the Vasicek model that used the Ornstein-Uhlenbeck (OU) process to model the dynamics of the short rates.

In the previous post, we built a finetuned VGG16-based model (Model 4) to detect malaria from cell images.

We continue our refresher series on the short-rate models. In the previous post, I introduced the Merton model and the Euler-Maruyama method to simulate it.

In the previous post, we replicated the Smooth Transition Exponential Smoothing (STES) model from (Taylor 2004) and (Liu, Taylor, Choo 2020).

I was honestly amazed by these results. Even the base model achieved 98% accuracy! However, accuracy isn’t everything. Let’s break down what these numbers really mean.

In the previous post of this series, we introduced the problem of Malaria detection and prepared and preprocessed our dataset.

I recently learned more about convolutional neural networks (CNNs) and how they are used to analyze images.

Let us continue our refresher series on short-rate models. We introduced the Merton short rate model in the previous post.

Welcome to my refresher series on short-rate models, dear reader! These mathematical marvels are not just academic curiosities but crucial tools in finance.

A value strategy posits that the worth of an asset tends to oscillate around a more gradual, underlying trend.

Exponential smoothing (ES) is a simple yet widely used approach to volatility forecasting in finance and economics.