Liquid Neural Network Models for Natural Gas Spot Price Time-Series Forecasting
Yiqian Liu, Jiayi Niu, Adam Kelleher, Subhabrata Das ∗ · 2026
Read Liquid Neural Network Models for Natural Gas Spot Price Time-Series Forecasting by Yiqian Liu, Jiayi Niu, Adam Kelleher, Subhabrata Das ∗ (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-24788.txt.
Introduction
Liu, Niu, Kelleher and Das explore Liquid Neural Networks for short-horizon forecasting of the Henry Hub natural gas spot price. Accurate short-term gas price forecasting is hard. Volatility comes from seasonal demand, geopolitical shifts and changing money conditions. Nonlinear dynamics and frequent regime changes limit traditional time-series models. Liquid networks adapt on the fly to changing temporal patterns through dynamic internal state updates, suiting nonstationary price behavior. Better forecasts in volatile conditions aim to reduce uncertainty for energy trading and power markets.
Catalog record
| CREATOR | Yiqian Liu, Jiayi Niu, Adam Kelleher, Subhabrata Das ∗ |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 49151 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.24788
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Liquid Neural Network Models for Natural Gas Spot Price Time-Series Forecasting Yiqian Liu, Jiayi Niu, Adam Kelleher, Subhabrata Das ∗ Abstract Natural gas is undoubtedly an essential component of the global energy system. Accurate short-term forecasting of natural gas price is challenging due to pronounced volatility driven by seasonal demand patterns, geopolitical developments, and shifting macroeconomic conditions. The nonlinear dynamics and frequent regime changes can limit the effectiveness of traditional time-series models. In this study, we explore the use of Liquid Neural Networks (LNNs) for short-horizon forecasting of the Henry Hub spot price, a primary benchmark for pricing. LNNs are designed to adapt continuously to evolving temporal patterns through dynamic internal state updates, making them well suited for nonstationary price behavior. By improving forecast accuracy in volatile market conditions, this work aims to reduce uncertainty and enhance decision support across energy trading and power market applications. I Introduction I-A Background and Motivation Natural gas remains a cornerstone of the global energy matrix, serving as a critical bridge fuel in the transition toward low-carbon energy systems while providing baseload reliability for power generation [ 17 ] . In the United States, the Henry Hub spot price functions as the primary benchmark for futures, options, and physical trading [ 16 ] , making its predictability of paramount importance to utility operators, energy traders, and financial institutions alike. The natural gas market is characterized by pronounced volatility and nonstationary behavior driven by weather shocks, geopolitical developments, and supply disruptions [ 5 ] . Traditional econometric models such as ARIMA and GARCH offer statistical interpretability, but are limited by linear assumptions and stationarity conditions during periods of structural change [ 2 ] . Rolling-window OLS addresses some of this rigidity through adaptive re-estimation, yet remains constrained by its linear functional form and cannot capture the nonlinear dependencies and regime-dependent volatility clustering characteristic of natural gas markets. Deep learning approaches, particularly LSTM networks, improve on this by modeling nonlinear temporal dependencies, but their fixed internal dynamics limit responsiveness when market regimes shift rapidly [ 8 ] . Thus, many existing approaches rely on fixed or discretely updated parameters, limiting their ability to continuously adapt to evolving market conditions. This study examines whether adaptivity improves short-term forecast accuracy compared to rolling-window linear regression. I-B Research Scope and Contributions To address this gap, we formulate Henry Hub price forecasting as a nonstationary time-series prediction problem and investigate Liquid Neural Networks (LNNs) — recurrent architectures whose dynamic time constants allow internal states to adapt continuously to incoming information — as a candidate solution. We construct a ten-and-a-half-year daily dataset (January 2015–August 2025) integrating Henry Hub spot prices with financial, energy, and operational variables, and focus on next-day short-horizon forecasting rather than long-term price projection. This study makes three principal contributions. First, we conduct a systematic comparison of five liquid neural architectures — LTC, Strict CfC, Hybrid CfC, CT-LTC, and a standard LSTM benchmark — against a rolling-window linear regression baseline for short-horizon Henry Hub forecasting. Second, we employ a stratified expanding-window evaluation protocol with Moving Block Bootstrap uncertainty quantification to produce statistically reliable performance estimates that account for temporal autocorrelation in nonstationary forecast errors. Third, through controlled architectural comparison, we identify which design characteristics — specifically, input-conditioned timescale modulation versus fixed or calendar-based step sizing — drive performance differences under the high-frequency regime shifts that typify natural gas markets. II Related Work Forecasting energy prices has been widely studied using both statistical and machine learning approaches. Early econometric models, such as autoregressive and cointegration frameworks, effectively captured short-term correlations but struggled during periods of structural market change caused by supply disruptions or policy shifts. Later, neural models such as LSTM and GRU improved nonlinear pattern recognition but often at the expense of interpretability and computational stability. Several studies have proposed rolling-window regression frameworks to better handle time-varying relationships. Dbouk and Jamali introduced a rolling estimation approach for daily crude oil prices, where model parameters are re-estimated within a fixed-length moving window to adapt to recent market conditions while discarding outdated information [ 4 ] . Artemova et al. and Khan et al. similarly demonstrate that rolling-window linear regression provides a strong baseline for short-term forecasting of volatile energy commodities [ 1 , 14 ] , supporting the principle that rolling estimation enhances robustness and responsiveness — a design choice we adopt in our baseline model. The challenge of applying deep learning to volatile energy time-series has motivated a range of hybrid and adaptive architectures. Das et al. propose a PCA-PR-Seq2Seq-Adam-LSTM framework for power outage prediction, demonstrating that structured preprocessing and temporal architectures substantially improve robustness under abrupt distributional shifts analogous to Henry Hub regime changes [ 3 ] . Multi-headed recurrent networks tuned via meta-heuristic optimization have also been applied to energy price forecasting, reinforcing the broader insight that models adapting to evolving dynamics outperform those with fixed parameters [ 9 ] . The study most closely related to ours is Liu et al. [ 15 ] , who compare feedforward networks, support vector machines, random forests, and LSTM for multi-step Henry Hub forecasting. Their results establish LSTM as a strong neural baseline and highlight systematic accuracy degradation with increasing horizon. However, their recurrent dynamics remain governed by fixed time scales, which may limit responsiveness under abrupt regime shifts. Zaragoza et al. address adaptivity from a different angle, proposing a time-varying graph learning framework for electricity price forecasting in which inter-node dependencies evolve dynamically [ 27 ] , underscoring a broader trend toward adaptive modeling in energy markets. Building on this literature, we introduce Liquid Neural Networks — recurrent architectures designed to capture fast-changing temporal dependencies through adaptive internal dynamics with fewer parameters and improved stability compared to standard recurrent models [ 7 , 6 ] , making them particularly suitable for the volatile, regime-prone behavior of Henry Hub spot prices. III Dataset III-A Data We utilize Henry Hub Natural Gas Spot Prices [ 20 ] as the target variable. Since gas pricing is driven by a mix of supply, demand, and market conditions, we construct a predictor set combining fundamentals and market indicators. WTI crude oil spot prices [ 19 ] serve as a benchmark for broader energy market trends and cross-commodity substitution. U.S. Treasury yield curve rates [ 18 ] — specifically the 2-year, 5-year, 10-year, and 20-year maturities — capture macroeconomic and capital cost conditions relevant to natural gas investment and trading. The U.S. Dollar Index [ 13 ] reflects dollar strength, which influences the competitiveness of U.S. LNG exports. Equity-based measures, including the S&P Energy Index [ 12 ] and the Dow Jones U.S. Coal Index [ 10 ] , provide signals of sector-wide sentiment and coal–gas substitution effects. Firm-level production sentiment is proxied using daily stock prices of EQT [ 11 ] . Finally, U.S. nuclear generation data — represented by installed capacity, outage capacity (both in megawatts), and percentage outage — reflects shifts in electricity supply capacity where reductions in nuclear availability can increase residual demand for natural gas [ 21 ] . All datasets were merged by date over the period January 6, 2015 to August 29, 2025. Variables with a high proportion of missing observations — specifically Coal Volatility, USD