Probabilistic DAM Price Forecasting for ERCOT Houston Hub

Summary

Day-Ahead Market (DAM) price forecasting in ERCOT presents a challenging time-series problem. Electricity prices exhibit strong intraday and weekly structures, but these patterns are not stationary. Changes in demand, renewable generation, system conditions, and market behavior can alter both the level and volatility of prices over time.

This case study examines an AI-based forecasting workflow for the ERCOT Houston Hub, using historical market data from October through December 2025 and forecasting the final ten days of December, from December 22 – December 31, 2025. The analysis places particular emphasis on an important modeling question for electricity markets: how much historical data should be used to train a price forecasting model? A longer historical dataset provides more observations, but it may also introduce outdated market behavior that is less representative of current conditions. The study therefore evaluates historical price behavior across multiple years before selecting a relatively short three-month training window. It then examines the relationship between Houston Hub price and ERCOT-wide net load, incorporates market-published wind and solar forecasts, and uses an AI forecasting workflow to generate a probabilistic DAM price forecast.

This case study used QuantRisk’s AI forecasting platform QR AMPS (AI Market Prediction System). For more information on our AI electricity price forecasting solutions, including technical architecture, case studies, and live interactive forecast dashboards, please visit:
https://quantrisk.com/analytics-electricity-price-forecast

1. A Look at the ERCOT Day-Ahead Market

The Electric Reliability Council of Texas (ERCOT) operates the wholesale electricity market for the ERCOT Interconnection. Its Day-Ahead Market is a forward, financially binding market in which participants can procure and sell energy and ancillary services for the following operating day. ERCOT describes DAM energy settlements using Settlement Point Prices calculated from DAM Locational Marginal Prices for Resource Nodes, Load Zones, and Hubs.

For this case study, the forecasting target is the ERCOT Houston Hub DAM price, which ERCOT publishes as a Settlement Point Price in its Day-Ahead Market data. ERCOT publishes DAM price information through its market-information infrastructure, including hourly DAM LMPs and DAM Settlement Point Prices.

For forecasting purposes, the objective is not to reproduce the published price after the market clears. It is to generate a forward-looking prediction before the DAM outcome is known, using information available at the forecasting timestamp. This makes the forecasting problem inherently time-dependent. The model must be trained using historical information while preserving the chronological separation between the information available at forecast time and the price that subsequently occurs.

ERCOT publishes system-wide Short-Term Wind Power Forecast (STWPF) and Short-Term Photovoltaic Power Forecast (STPPF) on an hourly basis, with each update providing forward-looking information for the coming hours and days. The analysis in this study uses the latest published values available in the historical dataset, while the forecasting workflow uses the most recent values available at the time each forecast is generated. This distinction preserves the information set that would have been available in an actual forecasting process and avoids using subsequently revised forecast values.

2. Forecast Inputs and Output

The forecasting workflow combines the historical Houston Hub price with market-published renewable-generation forecasts.

2.1 Input Data

The primary input datasets used in this study are:

• Historical ERCOT Houston Hub DAM price
• ERCOT Total Short-Term Wind Power Forecast (STWPF)
• ERCOT Total Short-Term Photovoltaic Power Forecast (STPPF)
• ERCOT Total demand-related information
• Derived time-series and calendar features

These inputs are particularly relevant to Houston Hub LMP because Houston prices are influenced not only by local demand, but also by the system-wide balance between ERCOT electricity demand and available generation. Wind and solar generation contribute to meeting system demand and therefore reduce the amount of conventional generation needed to supply the ERCOT system. Changes in this system-wide balance can affect the marginal generation resources setting energy prices, which in turn influences Houston Hub LMP, subject to local transmission constraints and congestion. For this reason, the forecasting model considers total demand together with expected wind and solar generation to capture the broader ERCOT market conditions affecting Houston prices.

2.2 Forecast Output

The output is a probabilistic forecast of Houston Hub DAM price.

Rather than producing only a single point forecast for each hour, the system generates a distribution of forecast outcomes. This allows the resulting forecast to communicate both the central prediction and the range of likely price outcomes.

3. Data Analysis: How Much Historical Data Should Be Used?

One of the central questions in this study was the appropriate length of the training dataset. A common assumption in time-series forecasting is that more historical data should improve the model because it provides more observations. For electricity prices, however, this assumption does not necessarily hold. Power markets evolve.

Generation portfolios change, renewable penetration changes, demand behavior changes, transmission conditions change, and market participants respond to changing economic and operational conditions. As a result, observations from several years ago may describe a market regime that is less representative of current conditions. For this reason, the study first examined approximately three years of historical Houston Hub price data to determine whether seasonal behavior remained sufficiently consistent across years.

3.1 Monthly Similarity Across Years

Monthly behavior can vary substantially even when two months have similar average price levels. A comparable monthly average does not necessarily imply comparable market behavior, as the underlying price volatility, frequency of spikes, and magnitude of price movements can differ significantly. As illustrated in the plot below, August exhibits higher price volatility despite having a similar average price to several other months, while February, May, and July show a greater number of pronounced price spikes and outliers.

The analysis further shows that the same calendar month can exhibit markedly different patterns across different years, as reflected in the low correlations between corresponding months in the historical heatmap.

The heatmap shows that most of corresponding months across different years do not exhibit particularly strong correlation. One of the more notable relationships is the correlation between September 2023 and September 2024, at approximately 59%. This suggests that seasonal labels alone do not guarantee comparable price behavior from one year to another.

These findings highlight the need to use enough historical data to capture a representative range of price volatility and spike behavior, while avoiding excessive reliance on older observations that may no longer reflect the current market regime. The training window must therefore balance historical coverage with the evolving nature of electricity-market behavior.

3.2 December Across Different Years

Since the forecast horizon in this study covers December 22 – 31, the analysis examines December price behavior across different years in greater detail to assess whether consistent patterns exist across years.

The December price profiles from 2023, 2024, and 2025 are not sufficiently similar to justify assuming that older Decembers provide a reliable template for December 2025. This reinforces the conclusion from the correlation analysis. For electricity price forecasting, recency can be more valuable than historical volume when market behavior changes over time.

Based on this analysis, the study selected a relatively short three-month training window, covering October through December 2025, with the final ten days of December reserved for forecasting evaluation. The forecasting period, December 22 – 31, 2025, is therefore treated as the out-of-sample period rather than being included in the training data.

4. Intraday and Weekly Price Structure

After selecting the training horizon, the next step is to examine the temporal structure of Houston Hub DAM prices within the selected data window.

4.1 Average Hourly Behavior

The average hourly profile shows a clear daily pattern.

Two morning hours, 06:00 and 07:00, show a distinct increase in price, followed by a more pronounced evening peak around 18:00. The stronger evening peak indicates that price levels vary materially across operating hours and reinforces the importance of including hour-of-day information as an input feature.

However, the average hourly curve reflects typical market behavior rather than the expected price for a specific operating hour. Houston Hub prices at 06:00, 07:00, or 18:00 can vary substantially from day to day depending on demand, renewable generation, and other prevailing market conditions. These relationships are examined through the feature analysis that follows.

4.2 Weekday-Hourly Behavior

The weekday analysis shows broadly similar intraday shapes across the week, but with meaningful differences in price levels.

The morning peaks are generally more pronounced on Mondays and Tuesdays, while Sundays exhibit the lowest overall levels. The evening profile shows another distinction: Saturday has the lowest evening peak, while Sunday follows a somewhat different pattern from the other days of the week and maintains a lower price level through approximately 15:00.

These differences indicate that calendar information provides useful context beyond the hour-of-day feature alone. The model therefore benefits from distinguishing between:

• Hour of day
• Day of week
• Weekday versus weekend
• Historical price behavior

The exploratory analysis also shows why these features should not be interpreted independently. The same operating hour can correspond to materially different price levels depending on the day and the prevailing system conditions.

5. Feature Analysis: From Houston Demand to ERCOT Total Net Load

The next stage of the analysis investigates the relationship between Houston Hub DAM LMP and ERCOT system conditions. A primary variable of interest is electricity demand.

Why ERCOT-Wide Net Load for a Houston Hub Forecast?

A natural question is why an ERCOT-wide net-load measure is relevant when the target is the Houston Hub price.

The reason is that Houston Hub prices are formed within the broader ERCOT market. System-wide supply-demand conditions influence the marginal generation resources available to serve demand across the ERCOT network. Although local transmission constraints and congestion can cause Houston prices to diverge from broader system conditions, the system-wide energy balance remains an important fundamental driver.

Therefore, for this study, ERCOT-wide net load is used as a system-level market fundamental, while the Houston Hub price remains the forecast target.

5.1 Total ERCOT Demand and Houston Hub Price

The first analysis compares the average Houston Hub LMP trend with different levels of total ERCOT demand.

The relationship is not consistently increasing across the full demand range. This is an important observation because it indicates that total demand alone does not fully characterize the market condition relevant to Houston Hub pricing. A higher system demand level does not necessarily translate directly into higher Houston Hub prices if a significant portion of that demand is being served by renewable generation. This motivates a more market-oriented transformation.

5.2 Net Demand: Accounting for Wind and Solar

The analysis therefore defines ERCOT net demand as the portion of system demand remaining after accounting for wind and solar generation:

ERCOT Net Demand = Total ERCOT Demand − System Wind Generation − System Solar Generation

When the analysis is repeated using net demand, the relationship with Houston Hub price becomes substantially clearer.

The resulting trend shows a clearer increase in average Houston Hub price as ERCOT net demand rises. This is consistent with the economic interpretation of net load. Demand represents the electricity that must be served, while wind and solar generation reduce the amount that must be supplied by the remaining generation fleet. Net demand therefore provides a more direct representation of the portion of system requirements that must be met after accounting for these renewable resources.

The below visualizations show how movements in ERCOT net load correspond with changes in Houston Hub price behavior. This type of analysis is particularly relevant for electricity-price forecasting because price volatility is often more difficult to model than the underlying average level.

5.3 Another Evidence for a Shorter Training Window

The study compares the correlation between Houston Hub LMP and ERCOT Total Net Load using different historical windows. Using approximately one year of data produces a Pearson correlation of about 62%. When the analysis is restricted to the selected three-month period, the correlation increases to approximately 77%. This difference provides additional evidence that the relationship between market fundamentals and price is not stable over long historical periods. The shorter window is more closely aligned with the market conditions represented in the forecasting period.

This does not mean that a three-month window will always be preferable for ERCOT price forecasting. Rather, the result demonstrates why training-window selection should be treated as a model-design decision and tested empirically, rather than assuming that the largest available historical dataset will always produce the best forecast.

6. No-Code Forecasting Workflow: From Data to Scheduled AI Forecast

The analytical process described above is implemented through the QR AMPS forecasting workflow, which provides a no-code, on-screen environment for data preparation, feature analysis, model configuration, and automated forecasting.

A key capability is the on-screen feature engineering and selection toolbox. It allows multiple datasets to be integrated and candidate forecasting inputs to be evaluated within the same environment. In this case, ERCOT-published market data, including price, demand, STWPF, and STPPF, can be brought into the workflow and transformed into model-ready features.

Data scientists can explore, transform, and evaluate different features within minutes without writing Python code. This provides a direct workflow from:

Market Data → Feature Engineering → Feature Selection → Model Inputs

The same environment also allows the modeling workflow to be configured without moving between disconnected analytical tools. Once the features are prepared, candidate Machine Learning and Deep Learning algorithms can be configured and evaluated through the platform.

6.1 Automated Forecast Scheduling

After model selection, the forecasting workflow should be scheduled for repeated execution. With QR AMPS, the schedule can be configured according to the desired forecasting process, while the forecast horizon can also be adjusted to the use case.

This allows the same workflow to support:

• Next-day forecasting
• Multi-day forecasting
• Longer forward horizons
• User-defined forecast windows

The important distinction is that the QuantRisk forecasting system is not limited to a one-time analytical exercise. The same configured workflow can repeatedly ingest new market information, generate the required features, execute the selected model, and produce the next forecast.

6.2 Probabilistic Houston Hub Forecast

A single point forecast provides one predicted outcome. A probabilistic forecast provides a broader view of the expected price distribution and therefore provides more information for risk-sensitive decisions. The QR AMPS final model produces a probabilistic DAM price forecast rather than only a single point prediction.

The forecast dashboard provides a distribution of possible price outcomes for each forecast hour. Where supported by the configured workflow, the system can generate 100 price forecast percentiles within minutes, providing a high-resolution representation of the predicted price distribution. This allows users to examine:

• Central forecast levels
• Lower and upper forecast ranges
• Potential high-price outcomes
• Potential low-price outcomes

The likelihood of outcome variation across forecast hours

The probabilistic representation is particularly relevant to electricity markets because price behavior can change rapidly when system conditions move away from typical operating ranges.

7. Forecast Results and Performance Analysis

The forecasting model is evaluated for the last ten days of the year 2025: December 22 – 31, 2025. The comparison between forecast and observed prices provides the first direct assessment of how well the model captured the realized market behavior.

The objective is not simply to minimize average error. A useful forecasting system should also capture the direction and structure of market movements.

7.1 Forecast Error Metrics

The forecasting dashboard evaluates performance using multiple complementary metrics:

7.2 Directional Win Rate

The directional win rate measures how frequently the model correctly predicts the direction of price movement according to the defined methodology. This provides an additional measure of model behavior that complements the numerical error metrics. Directional performance is particularly relevant when the forecast is used to support decisions where the primary question is whether prices are expected to move higher or lower.

MAPE and MAE quantify the magnitude of forecast errors, while directional win rate evaluates whether the model correctly anticipates the direction of market movement. Together with RMSE, these metrics provide a broader evaluation of forecasting performance.

8. Conclusion

The ERCOT Houston Hub study demonstrates an important principle in electricity-price forecasting: model design should reflect the changing nature of the market rather than simply maximize the amount of historical data available.

The analysis of three years of Houston Hub price behavior showed that corresponding months across different years were not consistently similar. The differences observed across December 2023, December 2024, and December 2025 further supported the use of a shorter, more recent training window. Within the selected three-month period, the data revealed clear intraday and weekly price structures. Morning peaks around 06:00–07:00 and a stronger evening peak around 18:00 were evident, while weekday analysis showed additional differences in both morning and evening behavior.

Feature analysis provided another important result. Total ERCOT demand did not exhibit a sufficiently consistent relationship with Houston Hub price on its own. After accounting for wind and solar generation, however, ERCOT net load provided a substantially clearer relationship with price. The increase in correlation from approximately 62% using one year of data to 77% using the selected three-month window further demonstrated the importance of aligning the training period with the current market regime.

The resulting forecasting workflow combines these market observations with QR AMPS no-code, on-screen environment for feature engineering, feature selection, model configuration, algorithm testing, backtesting, and automated scheduling. This creates a reproducible path from market-published data to a production forecasting workflow without requiring each analytical step to be implemented manually in code.

The final output is probabilistic rather than limited to a single point prediction. By producing multiple forecast percentiles, the system provides a more detailed representation of the expected price distribution. Performance is then assessed through MAPE, MAE, RMSE, and directional win rate, providing complementary views of numerical accuracy and market-direction prediction.

If you have any comments or questions, please contact us at [email protected]. Our team will be delighted to assist you and discuss your forecasting requirements.