🤖 machine learning

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

This paper proposes using Liquid Neural Networks to improve the short-term forecasting accuracy of Henry Hub natural gas spot prices by leveraging their ability to adapt continuously to the market's volatile, nonstationary, and regime-shifting dynamics.

Original authors: Yiqian Liu, Jiayi Niu, Adam Kelleher, Subhabrata Das

Published 2026-04-29

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Yiqian Liu, Jiayi Niu, Adam Kelleher, Subhabrata Das

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the price of natural gas as a wild, unpredictable horse. Sometimes it gallops gently, but often it spooks, jumps, or runs in circles due to sudden storms, geopolitical arguments, or changes in the weather. Trying to predict where this horse will go tomorrow is incredibly hard because it doesn't follow a straight line; it changes its mind constantly.

This paper is a race between different "trainers" (computer models) to see which one can best predict the horse's next move. The researchers from Columbia University wanted to see if a new type of AI, called a Liquid Neural Network (LNN), could do a better job than the old-school methods.

Here is the breakdown of their experiment and what they found, using simple analogies:

The Contestants

The researchers set up a race with five different types of models to predict the daily price changes of natural gas (specifically at the Henry Hub, which is like the "standard price tag" for gas in the US).

The Old Coach (Rolling-Window Linear Regression): This is the traditional method. Imagine a coach who looks at the last 30 days of the horse's behavior and draws a straight line to guess tomorrow. It's simple, but it assumes the horse will keep running in a straight line, which is rarely true in the energy market.
The Rigid Robot (LSTM): This is a standard AI that remembers the past. Think of it as a robot that has a fixed internal clock. It learns patterns, but once it sets its "gears," they don't change speed. If the market suddenly shifts gears, the robot is slow to react because its internal clock is stuck at a fixed pace.
The Liquid Models (LNNs): These are the new kids on the block. Instead of having a fixed clock, imagine these models have liquid time. Their internal gears can speed up or slow down instantly depending on what's happening right now. If the market is calm, they move slowly to save energy; if the market is chaotic, they speed up to catch every detail.

The researchers tested five specific versions of these "liquid" models:

Strict CfC: A liquid model with a fixed mixing rule.
LTC (Liquid Time-Constant): A model that can adjust its speed but uses a uniform step size.
Hybrid CfC: A model that adjusts its speed based on how "loud" the market noise is.
CT-LTC: A model that tries to adjust its speed based on the calendar (e.g., moving slower over weekends).

The Race Conditions

The data covered over 10 years (2015–2025), including massive shocks like the 2021 Winter Storm Uri (when gas prices went crazy) and the 2022 energy crisis. The models had to predict the price change for just one day ahead.

To make it a fair test, they didn't just give the models all the data at once. They used a "growing window" approach:

The models started with 30 days of history.
They made a prediction.
Then, they added one more day of real data to their history and tried again.
This repeated 160 times, simulating a real trader learning day by day.

The Results: Who Won?

1. The Old Coach (Linear Regression) came in last.
As expected, the simple straight-line model failed miserably. It couldn't handle the wild jumps and turns of the gas market. It was like trying to predict a rollercoaster ride by drawing a straight line.

2. The Rigid Robot (LSTM) was okay, but not great.
The standard AI did better than the Old Coach, but it struggled when the market suddenly changed its personality. Because its internal clock was fixed, it couldn't adapt quickly enough to the sudden "shocks" (like a storm hitting Texas). It tended to smooth over the wild spikes, missing the big moves.

3. The Liquid Models (LNNs) took the lead.
The new liquid models generally outperformed the others. They were better at handling the chaos.

The Winner: Hybrid CfC.
This model was the champion. Its secret sauce was input-conditioned timescale modulation.
- The Analogy: Imagine driving a car. When the road is smooth, you cruise at a steady speed. But when you see a pothole or a deer jump out, you instantly slam on the brakes or swerve. The Hybrid CfC does this: when the market data gets "loud" or chaotic, it instantly speeds up its internal processing to react. When things are calm, it slows down. This ability to change its "reaction speed" based on the current situation made it the most accurate.
The Runner-Up: LTC.
This model was also very strong. It used a continuous-time approach that allowed it to track the market's flow better than the rigid robots. However, it was slightly less flexible than the Hybrid CfC because it didn't adjust its speed as dynamically based on the specific "loudness" of the data.
The Disappointment: CT-LTC.
This model tried to be smart by looking at the calendar (e.g., knowing that Monday is three days after Friday). The researchers thought this would help, but it didn't work well.
- Why? Because in the gas market, a quiet Monday is very different from a Monday after a massive storm. The model treated all Mondays the same way just because of the calendar, missing the actual market news. It was like a driver who only looks at the clock but ignores the traffic ahead.

The Big Takeaway

The paper concludes that flexibility is key.

In a market that changes its mind constantly (non-stationary), models with fixed internal clocks (like the standard LSTM) struggle. The best performers were the ones that could continuously adapt their internal speed to match the chaos of the moment.

The Hybrid CfC proved that the best way to predict a wild, volatile market isn't to have a faster computer, but to have a computer that knows when to speed up and when to slow down based on what is happening right now.

What the paper didn't say:

It did not claim these models can predict the future perfectly (they still missed the biggest, most extreme shocks like Winter Storm Uri).
It did not suggest using these models for long-term planning (like 5 years out); they are designed for short-term, next-day guesses.
It noted that the models are currently missing some "physical" data (like actual pipeline flows or weather forecasts), which limits how well they can predict the biggest jumps.

In short: To predict a wild horse, you need a trainer who can change their strategy instantly, not one who follows a rigid schedule. The Hybrid Liquid Neural Network is that flexible trainer.

1. Problem Statement

The paper addresses the challenge of short-term forecasting of natural gas spot prices (specifically the Henry Hub benchmark). This task is inherently difficult due to:

High Volatility and Non-stationarity: Prices are driven by seasonal demand, geopolitical shocks, and supply disruptions, leading to frequent regime changes.
Limitations of Traditional Models:
- Econometric models (ARIMA, GARCH): Rely on linear assumptions and stationarity, failing during structural breaks.
- Standard Deep Learning (LSTM): While capable of modeling non-linearities, they use fixed internal dynamics and discrete time steps, limiting their ability to adapt continuously to rapid market shifts.
The Gap: Existing approaches often rely on fixed parameters or discrete updates, lacking the continuous adaptability required for volatile commodity markets.

2. Methodology

A. Dataset

Target: Daily Henry Hub Natural Gas Spot Prices (USD/MMBtu).
Period: January 6, 2015 – August 29, 2025 (2,645 observations).
Features: 30 predictor variables including:
- Energy Fundamentals: WTI Crude Oil, Nuclear generation capacity/outages.
- Macroeconomics: U.S. Treasury yield curve (2y, 5y, 10y, 20y), U.S. Dollar Index.
- Market Sentiment: S&P Energy Index, Dow Jones Coal Index, EQT Corporation stock prices.
Preprocessing: Missing data (holidays) and extreme outliers (beyond 5.5 std dev) were handled carefully to avoid survivorship bias while retaining genuine market shocks (e.g., Winter Storm Uri).

B. Baseline Model

Multivariate Rolling-Window Linear Regression: A standard benchmark using a 30-day rolling window to re-estimate coefficients daily. It includes lagged returns and exogenous features.

C. Neural Architectures Evaluated

The study compares five neural models against the baseline:

LSTM: Standard discrete-time recurrent benchmark.
Strict CfC (Closed-form Continuous-time): Uses a closed-form solution with a sigmoid interpolation gate but lacks input-dependent time constants.
LTC (Liquid Time-Constant): Uses continuous-time ODEs with learnable time constants ( $\tau$ ) and attractor parameters, updated via semi-implicit Euler steps.
Hybrid CfC: Combines closed-form updates with input-conditioned time-scale modulation. The decay rate adapts based on the magnitude of the input signal ( $|f_t|$ ).
CT-LTC (Calendar-gap LTC): An extension of LTC where the integration step size ( $\Delta t$ ) is determined by the actual calendar time elapsed (accounting for weekends/holidays) rather than a fixed unit step.

D. Evaluation Protocol

Expanding Window: Models are retrained from scratch at each evaluation point using an expanding history (unlike rolling windows) to leverage growing data, which benefits LNNs.
Stratified Sampling: 160 evaluation points drawn from 20 temporal bins to ensure coverage across different market regimes.
Uncertainty Quantification: Moving Block Bootstrap (MBB) with 300 replications is used to generate 95% confidence intervals for metrics, preserving temporal autocorrelation in forecast errors.
Metrics: Pearson correlation ( $r$ ), Spearman rank ( $\rho$ ), Directional Accuracy (DA), $R^2$ , RMSE, and MAE.

3. Key Contributions

Systematic Comparison: First comprehensive evaluation of five distinct Liquid Neural Network (LNN) architectures against a robust linear baseline for natural gas forecasting.
Rigorous Statistical Validation: Implementation of a stratified expanding-window protocol combined with Moving Block Bootstrap to provide statistically reliable performance estimates that account for non-stationarity and temporal autocorrelation.
Architectural Insight: Identification of input-conditioned timescale modulation as the critical design feature driving performance in high-frequency, regime-shifting markets, distinguishing it from fixed-step or calendar-based approaches.

4. Results and Analysis

Performance Summary

Baseline (Linear Regression): Performed poorly ( $r \approx -0.04$ , $R^2 \approx -91$ ), confirming the non-linear nature of the data.
LSTM: Achieved moderate directional accuracy (55%) but failed to capture magnitude ( $R^2 \approx 0$ ), indicating it smooths over critical volatility.
Strict CfC: Underperformed LSTM, suggesting that closed-form updates without adaptive time constants are insufficient for this task.
CT-LTC: Despite theoretical appeal regarding calendar gaps, it underperformed the uniform-step LTC. The authors argue that calendar distance is a poor proxy for "informational distance" in event-driven markets.
LTC: Significantly outperformed LSTM and Strict CfC ( $r \approx 0.25$ , positive $R^2$ ). Its continuous-time dynamics allowed better tracking of return path levels.
Hybrid CfC (Best Performer):
- Highest Metrics: Achieved the highest Pearson correlation ( $r \approx 0.30$ ), Spearman correlation, and $R^2$ (0.057).
- Statistical Significance: Bootstrap confidence intervals for $R^2$ were strictly positive, confirming the advantage is systematic, not random.
- Mechanism: Its success is attributed to the adaptive decay rate ( $A_t = 1/\tau_t + |f_t|$ ). When market shocks occur (large $|f_t|$ ), the model rapidly sheds stale information; during quiet periods, it integrates over longer horizons.

Key Findings

Directional vs. Magnitude: While some models (like LSTM) could predict direction slightly better than random, only Hybrid CfC and LTC successfully predicted the magnitude and structure of returns.
Adaptivity is Key: The ability to modulate internal time constants based on input magnitude (Hybrid CfC) is superior to fixed time constants (LTC) or calendar-based steps (CT-LTC) for volatile energy markets.
Extreme Events: All models struggled to predict extreme exogenous shocks (e.g., Winter Storm Uri), highlighting the limitation of purely data-driven models without real-time external signals (e.g., weather forecasts).

5. Significance and Conclusion

Practical Impact: The study demonstrates that Hybrid CfC and LTC provide a reliable framework for short-term natural gas forecasting. Improved accuracy can enhance decision-making for utilities, traders, and risk managers, particularly during periods of high volatility.
Theoretical Contribution: The paper validates that integrating adaptive temporal mechanisms into neural architectures is a principled approach for modeling complex financial time series where stationarity assumptions fail.
Future Directions: The authors suggest future work should focus on:
- Incorporating physical market fundamentals (storage reports, pipeline flows, weather data) rather than just financial proxies.
- Extending the framework to multi-step forecasting horizons relevant for hedging.
- Comparing against classical time-series benchmarks (ARIMA/GARCH) to isolate the specific gains from non-linearity vs. adaptive re-estimation.

In summary, the paper establishes that Liquid Neural Networks, specifically the Hybrid CfC architecture, outperform traditional linear and standard deep learning models in forecasting natural gas prices by effectively adapting their internal dynamics to the non-stationary and shock-driven nature of energy markets.