pith. sign in

arxiv: 2604.24788 · v1 · submitted 2026-04-24 · 💻 cs.LG · cs.AI

Liquid Neural Network Models for Natural Gas Spot Price Time-Series Forecasting

Yiqian Liu , Jiayi Niu , Adam Kelleher , Subhabrata Das This is my paper

Pith reviewed 2026-05-08 12:04 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords liquid neural networkstime series forecastingnatural gas pricesHenry Hubnonstationary time seriesenergy price forecastingvolatile markets
0
0 comments X

The pith

Liquid neural networks forecast natural gas spot prices by continuously adapting their internal states to volatile, regime-shifting data.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper explores liquid neural networks for short-term forecasting of the Henry Hub natural gas spot price. These networks maintain dynamic internal states that update continuously in response to new price observations. The authors argue that this mechanism handles the nonlinear dynamics and frequent regime changes caused by seasonal demand, geopolitics, and macroeconomic shifts better than fixed traditional models. If the approach works, it would supply more reliable short-horizon predictions for energy trading and power market operations.

Core claim

Liquid Neural Networks are well suited for short-horizon forecasting of the Henry Hub spot price because their continuous adaptation through dynamic internal state updates allows them to capture nonlinear dynamics and frequent regime changes in nonstationary energy price series.

What carries the argument

Liquid Neural Networks: recurrent architectures whose internal state evolves continuously over time to match evolving input patterns.

If this is right

  • Short-term price forecasts become more accurate under volatile market conditions.
  • Energy traders and power market participants gain reduced uncertainty in decision making.
  • Continuous state updates enable the model to track shifts without retraining from scratch.
  • The same adaptation property extends to other nonstationary price series in energy markets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • LNNs could be tested on other volatile commodities such as oil or electricity prices to check generalization.
  • Pairing the model with live data streams might support intraday trading adjustments.
  • Explicit comparisons to LSTM and transformer baselines on identical data splits would quantify the adaptation advantage.

Load-bearing premise

The nonlinear dynamics and frequent regime changes in natural gas prices can be effectively captured and forecasted by the continuous adaptation mechanism of liquid neural networks.

What would settle it

A side-by-side evaluation on held-out Henry Hub data in which liquid neural network forecasts show no accuracy gain over standard models such as ARIMA or LSTM would disprove the claimed suitability.

Figures

Figures reproduced from arXiv: 2604.24788 by Adam Kelleher, Jiayi Niu, Subhabrata Das, Yiqian Liu.

Figure 1
Figure 1. Figure 1: Henry Hub Natural Gas Spot Price Trend (2015–2025) view at source ↗
Figure 4
Figure 4. Figure 4: 30-Day Rolling Standard Deviation of Spot Returns view at source ↗
Figure 5
Figure 5. Figure 5: Autocorrelation Function of Natural Gas Spot Price Returns view at source ↗
Figure 6
Figure 6. Figure 6: Baseline 30-Day Rolling Regression: Actual vs. Predicted Spot view at source ↗
Figure 7
Figure 7. Figure 7: Baseline 30-Day Rolling Regression: Actual vs. Predicted Spot view at source ↗
Figure 8
Figure 8. Figure 8: LSTM model: actual vs. predicted spot returns. view at source ↗
Figure 9
Figure 9. Figure 9: Strict CfC model: actual vs. predicted spot returns. view at source ↗
Figure 11
Figure 11. Figure 11: Hybrid CfC model: actual vs. predicted spot returns. view at source ↗
Figure 10
Figure 10. Figure 10: Liquid Time-Constant (LTC) model: actual vs. predicted spot view at source ↗
Figure 12
Figure 12. Figure 12: CT-LTC model: actual vs. predicted spot returns. view at source ↗
read the original 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The paper claims that Liquid Neural Networks (LNNs) are well suited for short-horizon forecasting of the Henry Hub natural gas spot price because their continuous internal state updates enable adaptation to nonlinear dynamics, regime changes, and nonstationary behavior in volatile price series, thereby improving accuracy over traditional models and reducing uncertainty for energy trading applications.

Significance. If the central claim were supported by rigorous experiments, the work could provide a useful modeling approach for handling volatility in a key energy commodity, with potential benefits for decision support in trading and power markets. The focus on dynamic adaptation aligns with known difficulties in commodity price forecasting.

major comments (1)
  1. [Abstract] Abstract: The manuscript asserts that LNNs improve forecast accuracy in volatile conditions due to dynamic state updates, yet supplies no dataset description, training procedure, loss function, evaluation metrics (e.g., RMSE, MAE), baseline comparisons (ARIMA, GARCH, LSTM), or any empirical results. The central claim therefore remains an untested hypothesis rather than a demonstrated result.

Simulated Author's Rebuttal

1 responses · 0 unresolved

Thank you for the referee's feedback. We acknowledge the validity of the concern regarding the absence of empirical details and will revise the manuscript accordingly to include a full experimental evaluation.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The manuscript asserts that LNNs improve forecast accuracy in volatile conditions due to dynamic state updates, yet supplies no dataset description, training procedure, loss function, evaluation metrics (e.g., RMSE, MAE), baseline comparisons (ARIMA, GARCH, LSTM), or any empirical results. The central claim therefore remains an untested hypothesis rather than a demonstrated result.

    Authors: We agree that the current manuscript does not supply the requested details on the dataset, training procedure, loss function, evaluation metrics, baseline comparisons, or empirical results, rendering the central claim unsupported in its present form. This is a clear limitation of the submitted version. In the revised manuscript we will add a dedicated Experiments section that provides: a description of the Henry Hub daily spot price dataset (source, time span, preprocessing); the LNN training procedure and hyperparameters; the loss function; the metrics (RMSE, MAE, and others); explicit comparisons to ARIMA, GARCH, LSTM and related baselines; and the corresponding numerical results with tables and figures. These additions will convert the claim from an untested hypothesis into an empirically evaluated result. revision: yes

Circularity Check

0 steps flagged

No derivation chain or self-referential predictions present

full rationale

The manuscript contains no equations, derivations, fitted parameters, or claimed first-principles results. It asserts that LNNs are well suited to nonstationary natural gas prices due to their dynamic internal state updates, but this is presented as a known architectural property rather than a derived claim that reduces to the paper's own inputs. No self-citations, ansatzes, or uniqueness theorems are invoked to support a load-bearing step. The work is an exploratory proposal without any mathematical or empirical reduction that could be circular.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; the central claim rests on the untested assumption that LNN adaptation will handle price volatility.

pith-pipeline@v0.9.0 · 5427 in / 998 out tokens · 39545 ms · 2026-05-08T12:04:41.149813+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

46 extracted references · 3 canonical work pages

  1. [1]

    Henry Hub Natural Gas Spot Price Dynamics:Figure 1 presents the daily Henry Hub natural gas spot price together with a 30-day moving average, with observations exceeding the 95th percentile highlighted to emphasize periods of ex- treme price behavior. The series exhibits pronounced volatility and distinct regime changes, with particularly large price disl...

    2021

  2. [2]

    Compared to other series, natural gas cumulative returns (black line) exhibit extreme volatility, ranging from -80% to +220%, while the USD Index (purple) barely moves beyond ±15%

    Energy and Financial Market Performance:Figure 2 places Henry Hub in the energy and finance context. Compared to other series, natural gas cumulative returns (black line) exhibit extreme volatility, ranging from -80% to +220%, while the USD Index (purple) barely moves beyond ±15%. Several structural observations stand out. Coal (orange dashed) has lost 60...

    2015

  3. [3]

    The heatmap shows weak linear correlations between the spot price and most individual predictors, with the exception of capacity (MW) and WTI crude oil spot prices

    Feature Correlations:Figure 3 shows the pairwise Pearson correlation structure between the Henry Hub spot price and selected predictor variables. The heatmap shows weak linear correlations between the spot price and most individual predictors, with the exception of capacity (MW) and WTI crude oil spot prices. In contrast, several predictors show strong in...

  4. [4]

    The plot demonstrates distinct volatility clustering, a phenomenon where periods of relative stability are interrupted by sudden, sustained regimes of high variance

    Rolling 30-Day Volatility:Figure 4 illustrates the re- alized volatility of natural gas spot returns, calculated as the 30-day rolling standard deviation. The plot demonstrates distinct volatility clustering, a phenomenon where periods of relative stability are interrupted by sudden, sustained regimes of high variance. This behavior provides strong eviden...

  5. [5]

    The ACF reveals two significant autocorrelation zones relative to the 95% confidence interval (±0.038)

    Autocorrelation Function:The ACF plot (Figure 5) demonstrates the Autocorrelation Function (ACF) of daily nat- ural gas spot returns, supporting a 30-day expanding window size for the LNN model. The ACF reveals two significant autocorrelation zones relative to the 95% confidence interval (±0.038). The first zone, spanning lags 1–8, has the most negative a...

  6. [6]

    Leth t andc t denote the hidden and cell states at timet

    LSTM:As a nonlinear recurrent benchmark, we use a standard single-layer LSTM. Leth t andc t denote the hidden and cell states at timet. The recurrence is ft =σ(W f xt +U f ht−1 +b f),(4) it =σ(W ixt +U iht−1 +b i),(5) ot =σ(W oxt +U oht−1 +b o),(6) ˜ct = tanh(Wcxt +U cht−1 +b c),(7) ct =f t ⊙c t−1 +i t ⊙ ˜ct,(8) ht =o t ⊙tanh(c t).(9) The final hidden sta...

  7. [7]

    Letu t = [xt;h t−1]denote the concatenation of the current input and previous hidden state

    Strict CfC:The Strict CfC implements the gated- interpolation form of the Closed-form Continuous-time Neural Network [14]. Letu t = [xt;h t−1]denote the concatenation of the current input and previous hidden state. A shared backbone MLP, given by one linear layer followed by Tanh, mapsu t to st = tanh(Wbbut +b bb)∈R m,(11) wherem= max(2n,32); in our exper...

  8. [8]

    LTC (uniform-step):The LTC model implements a Liquid Time-Constant network whose hidden state evolves ac- cording to the continuous-time dynamics of Hasaniet al.[13]. For hidden uniti, the underlying continuous-time dynamics are dxi(t) dt =− 1 τi +f i(x(t),I(t), θ) xi(t)+fi(x(t),I(t), θ)A i, (16) whereτ i >0is a learnable base time constant andA i is a le...

  9. [9]

    Hybrid CfC:The Hybrid CfC uses a closed-form re- current update with input-dependent time-scale modulation. At each step, the input and previous hidden state are concatenated as ut = [xt;h t−1].(22) A nonlinear drive and adaptive time constant are then com- puted by ft = tanh(Wf ut +b f),(23) τ t = exp(θτ)⊙σ(W τ ut +b τ),(24) with elementwise lower clippi...

  10. [10]

    The underlying ODE, log-space time constant parameterization, nonlinear response, and fused semi-implicit Euler update structure are identical to those of LTC

    CT-LTC (calendar∆t):CT-LTC extends the LTC formulation (Section IV-B3) by replacing the fixed unit step size with the observed calendar gap between consecutive timestamps. The underlying ODE, log-space time constant parameterization, nonlinear response, and fused semi-implicit Euler update structure are identical to those of LTC. The sole modification is ...

  11. [11]

    Extreme Event Behavior and Forecast Deviations:For all model predictions, visualizations comparing predicted and actual returns reveal several large and abrupt deviations that none of the evaluated models adequately capture, reflecting ex- treme market movements caused by exogenous shocks rather than typical market dynamics. One example occurred in Februa...

    2021

  12. [12]

    Baseline Model: AR(1) Spot Price + All Features: Using a 30-day rolling window, the baseline model generated Table II MODEL PERFORMANCE ON THE TEST SET AND UNDER BOOTSTRAP RESAMPLING. Panel A: Test-set performance Model PearsonrSpearmanρDA (%)R 2 RMSE MAE Baseline−0.0408−0.0122 43.30−91.03 61.12 32.45 LSTM0.1035 0.1714 55.00−0.0133 8.129 5.149 Strict CfC0...

  13. [13]

    LSTM:As reported in Table II, the LSTM achieves the highest directional accuracy among the non-hybrid ar- chitectures, yet itsR 2 remains near zero and its correlation coefficients are the weakest in the neural model tier. This dissociation reflects a structural limitation of the discrete gating mechanism: the forget, input, and output gates regulate info...

  14. [14]

    Strict CfC:As shown in Table II, the Strict CfC fails to outperform the LSTM in any metric of practical significance. Despite its continuous-time motivation, the architecture’s sig- moid interpolation gate controls only the mixture ratio between two candidate trajectory limits, not the rate at which the hidden state transitions between them. A large retur...

  15. [15]

    Pearson and Spearman correlations are nearly double those of weaker architectures, with positiveR 2 and low error metrics in the neural tier

    LTC:The LTC model greatly outperforms the LSTM and Strict CfC, as shown in Table II. Pearson and Spearman correlations are nearly double those of weaker architectures, with positiveR 2 and low error metrics in the neural tier. Significantly, these improvements focus on magnitude-related metrics, while directional accuracy remains modest. The LTC architect...

  16. [16]

    Hybrid CfC:The Hybrid CfC is the best overall performer across all metrics in both the point-estimate and bootstrap panels, as shown in Table II. Its superiority over the LTC — the second-strongest model — is narrow in absolute terms but consistent across all resampled evaluation windows, which is the more important criterion for assessing genuine archite...

  17. [17]

    CT-LTC:The CT-LTC model extends the LTC formu- lation by replacing the fixed unit step size with the observed calendar gap between consecutive timestamps, encoding week- end and holiday spacing directly into the integration dynamics. Despite the theoretical appeal of this modification — financial time series are inherently irregularly spaced, and Monday r...

  18. [18]

    A significantly nonzero mean error would indicate that a model consistently over- or under- predicts spot returns

    Forecast Bias:To assess whether the evaluated models produce systematically biased forecasts, we compute the mean forecast error for each architecture and test its significance using a two-sidedt-test. A significantly nonzero mean error would indicate that a model consistently over- or under- predicts spot returns. None of the five neural architectures ex...

  19. [19]

    Natural gas: A transition fuel for sustainable energy system transformation?

    A. Safari, N. Das, O. Langhelle, J. Roy, and M. Assadi, “Natural gas: A transition fuel for sustainable energy system transformation?”Energy Science & Engineering, vol. 7, no. 4, pp. 1075–1094, 2019

    2019

  20. [20]

    Henry hub and national balancing point prices: What will be the international gas price reference?

    A. E. H. Mazighi, “Henry hub and national balancing point prices: What will be the international gas price reference?”OPEC Review, vol. 29, no. 3, pp. 219–230, 2005

    2005

  21. [21]

    What drives volatility in natural gas prices?

    A. Hailemariam and R. Smyth, “What drives volatility in natural gas prices?”Energy Economics, vol. 80, pp. 731–742, 2019

    2019

  22. [22]

    Does natural gas consumption follow a nonlinear path over time? evidence from 50 U.S. states,

    A. Aslan, “Does natural gas consumption follow a nonlinear path over time? evidence from 50 U.S. states,”Renewable and Sustainable Energy Reviews, vol. 15, no. 9, pp. 4466–4472, 2011

    2011

  23. [24]

    Available: https://arxiv.org/abs/2006.10119

    [Online]. Available: https://arxiv.org/abs/2006.10119

    work page arXiv 2006

  24. [25]

    Predicting daily oil prices: Linear and non-linear models,

    W. Dbouk, “Predicting daily oil prices: Linear and non-linear models,” inProceedings of the 2018 EFMA Annual Meetings, Milan, Italy, 2018, eFMA Paper No. 0192, European Financial Management Association Annual Meeting

    2018

  25. [26]

    On the use of flexible rolling-window estimation and forecasting with evidence from the Great Recession to the COVID-19 pandemic,

    M. Artemova, F. Blasques, S. J. Koopman, and Z. Zhang, “On the use of flexible rolling-window estimation and forecasting with evidence from the Great Recession to the COVID-19 pandemic,” Working paper, Tinbergen Institute and Vrije Universiteit Amsterdam,

  26. [27]

    Available: https://www.tse-fr.eu/sites/default/files/TSE/ documents/sem2025/eee/artemova.pdf

    [Online]. Available: https://www.tse-fr.eu/sites/default/files/TSE/ documents/sem2025/eee/artemova.pdf

  27. [28]

    Sliding window regression based short-term load forecasting of a multi-area power system,

    I. A. Khan, A. Akbar, and Y . Xu, “Sliding window regression based short-term load forecasting of a multi-area power system,” arXiv preprint arXiv:1905.08111, 2019. [Online]. Available: https: //arxiv.org/abs/1905.08111

    work page arXiv 1905

  28. [29]

    A hybrid PCA-PR-Seq2Seq-Adam- LSTM framework for time-series power outage prediction,

    S. Das, B. Khan, and X.-Y . Liu, “A hybrid PCA-PR-Seq2Seq-Adam- LSTM framework for time-series power outage prediction,”arXiv preprint arXiv:2509.16743, 2025

    work page arXiv 2025

  29. [30]

    Gasoline and crude oil price prediction using multi- headed variational neighbour search-tuned recurrent neural networks,

    F. Ilhanet al., “Gasoline and crude oil price prediction using multi- headed variational neighbour search-tuned recurrent neural networks,” Computational Economics, 2025

    2025

  30. [31]

    Natural gas price prediction based on artificial intelligence models,

    X. Liu, M. Tang, Y . Feng, T. Fang, Z. Wang, and M. Song, “Natural gas price prediction based on artificial intelligence models,”PLOS ONE, vol. 20, no. 12, p. e0336582, 2025

    2025

  31. [32]

    Time-varying graph learning for electricity price forecasting,

    H. Zaragoza, A. B. Granados, J. M. Peña-González, G. B. Giannakis, and C. G. Brinton, “Time-varying graph learning for electricity price forecasting,” in2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. 1–5

    2024

  32. [33]

    Liquid time- constant networks,

    R. Hasani, M. Lechner, A. Amini, D. Rus, and R. Grosu, “Liquid time- constant networks,” inProceedings of the Thirty-Fifth AAAI Conference on Artificial Intelligence (AAAI-21), vol. 35, no. 9. AAAI Press, 2021, pp. 7657–7666

    2021

  33. [34]

    Closed-form continuous-time neural networks,

    R. Hasani, M. Lechner, A. Aminiet al., “Closed-form continuous-time neural networks,”Nature Machine Intelligence, vol. 4, pp. 992–1003, 2022

    2022

  34. [35]

    Natural gas spot price (henry hub),

    U.S. Energy Information Administration, “Natural gas spot price (henry hub),” https://www.eia.gov/opendata/browser/natural-gas/pri/fut

  35. [36]

    Cushing, ok WTI spot price FOB,

    U.S. Energy Information Administration, “Cushing, ok WTI spot price FOB,” https://www.eia.gov/dnav/pet/hist/RWTCD.htm

  36. [37]

    Daily treasury par yield curve rates,

    U.S. Department of the Treasury, “Daily treasury par yield curve rates,” https://home.treasury.gov/resource-center/data-chart-center/ interest-rates/TextView

  37. [38]

    U.s. dollar index historical data,

    Investing.com, “U.s. dollar index historical data,” https://www.investing. com/indices/usdollar-historical-data

  38. [39]

    S&P 500 energy historical data,

    Investing.com, “S&P 500 energy historical data,” https://www.investing. com/indices/s-p-500-energy-historical-data

  39. [40]

    Dow jones u.s. coal total stock market historical data,

    Investing.com, “Dow jones u.s. coal total stock market historical data,” https://www.investing.com/indices/dow-jones-u.s. -coal-total-stock-mar-historical-data

  40. [41]

    EQT corporation historical data,

    Investing.com, “EQT corporation historical data,” https://www.investing. com/equities/eqt-corporation-historical-data

  41. [42]

    U.s. nuclear outages,

    U.S. Energy Information Administration, “U.s. nuclear outages,” https: //www.eia.gov/nuclear/outages/

  42. [43]

    Winter storm uri caused widespread disruptions to u.s. energy supply and demand,

    U.S. Energy Information Administration, “Winter storm uri caused widespread disruptions to u.s. energy supply and demand,” https://www. eia.gov/todayinenergy/detail.php?id=50778, 2021

    2021

  43. [44]

    U.s. natural gas prices in- creased amid strong demand and lower production,

    U.S. Energy Information Administration, “U.s. natural gas prices in- creased amid strong demand and lower production,” https://www.eia. gov/todayinenergy/detail.php?id=55119, 2022

    2022

  44. [45]

    Winter storm elliott disrupts u.s. energy markets and drives price volatility,

    U.S. Energy Information Administration, “Winter storm elliott disrupts u.s. energy markets and drives price volatility,” https://www.eia.gov/ todayinenergy/detail.php?id=55139, 2022

    2022

  45. [46]

    U.s. natural gas market dynamics and lng export activity in 2024,

    U.S. Energy Information Administration, “U.s. natural gas market dynamics and lng export activity in 2024,” https://www.eia.gov/ todayinenergy/detail.php?id=61563, 2024

    2024

  46. [47]

    U.s. natural gas storage levels remain elevated entering winter 2024–2025,

    U.S. Energy Information Administration, “U.s. natural gas storage levels remain elevated entering winter 2024–2025,” https://www.eia.gov/ todayinenergy/detail.php?id=64004, 2024

    2024