CAISO DAM Price Forecast Case Study

1.The Case

The California Independent System Operator (CAISO) manages the operation of California’s high-voltage transmission grid and administers one of the world’s largest wholesale electricity markets. Through competitive market operations and efficient grid management, CAISO ensures reliable electricity delivery while supporting the integration of renewable energy resources.

CAISO determines electricity prices using Locational Marginal Prices (LMPs), which reflect the cost of producing and delivering electricity at specific locations, or nodes, across the transmission network. LMPs incorporate the effects of energy costs, transmission congestion, and system losses, providing transparent price signals for market participants.

CAISO supports trading across several major trading hubs, including TH_SP15, TH_NP15, and TH_ZP26, in both the 15-minute Real-Time Market (FMM) and the hourly Day-Ahead (DA) Market.

This case study presents the performance of an hourly day-ahead market (DAM) price forecasting model developed for the TH_SP15 Hub LMP. The forecasting model is powered by QR AMPS (AI Market Prediction System), QuantRisk’s AI-driven energy forecasting platform.

The CAISO Day-Ahead Market accepts bids and schedules up to seven days before the delivery date and closes on the day prior to delivery at approximately noon (Pacific Time). To provide market participants with sufficient time to develop their trading strategies, QR AMPS generates and publishes the day-ahead price forecast each day at 6:00 a.m. California time for the following trading day.

The forecast generated by QR AMPS serves as a core input in trading optimization systems such as QR TOPS (Trading Optimization & Portfolio System), enabling market participants to optimize bidding strategies, improve asset scheduling, and make data-driven trading decisions.

The data presented in this case study is sourced from QuantRisk’s forecasting portal where forecasts are published for multiple US and global power markets.

2.The Data

The hourly TH_SP15 Hub LMP DAM time series serves as the primary target variable for training the AI forecasting model. Accurate electricity price forecasting relies heavily on effective feature engineering. To improve forecast accuracy, robustness, and stability, the model incorporates a diverse set of demand-side, supply-side, and commodity market variables published by CAISO, together with proprietary forecasts generated by QuantRisk.

2.1 Input Data

For this case study, the AI forecasting model is trained using six months of hourly market data, covering the period from February to August 2026.

The key input features include:

• Actual market data: Historical TH_SP15 Hub LMP prices published by CAISO. These prices represent the target variable that the AI model is trained to forecast.
• Demand-side data: 7 DA electricity demand forecast published by the ISO, capturing expected system and regional demand that influences wholesale electricity prices.
• Supply-side data: A comprehensive set of supply-side indicators, including regional fuel prices, day-ahead solar and wind generation forecasts, and generation outage information covering hydroelectric, renewable, and thermal resources. These variables provide valuable insight into generation availability and the underlying supply conditions that influence market prices.

2.2 Output Data

The output of the AI forecasting model is an hourly day-ahead forecast of the TH_SP15 Hub LMP. The forecast is generated and published daily at 6:00 a.m. California time for the following trading day, providing market participants with actionable price intelligence ahead of the Day-Ahead Market deadline.

3.Data Exploration

The first step in developing an effective AI forecasting model is to perform a detailed exploration of the available data. This process helps identify quantifiable behavioral patterns, market dynamics, and underlying relationships that may not be immediately apparent from raw data.

To accomplish this, the QuantRisk data science team makes extensive use of the QR Feature Engineering and Data Analysis toolbox, delivering advanced data analysis and visualization techniques in a no coding environment, where time series and mathematical processors can be dragged and dropped in a visual interface to build long chains of data analysis. The underlying Python libraries automatically adjust to the flow and execute the calculations to extract meaningful insights from historical market data. See section 4 below. The following analysis presents the six-month hourly TH_SP15 Hub LMP dataset used for AI model training.

The historical price profile reveals clear seasonal changes throughout the training period. These variations are primarily driven by weather and solar conditions, changes in electricity demand, and fluctuations in regional natural gas prices resulting from pipeline constraints. A more granular analysis is therefore required to identify the most informative predictive features and optimize model performance.

3.1 Hourly Price Pattern Analysis

The following analysis presents the average hourly TH_SP15 Hub LMP across the six-month training period, from February to August 2026. This analysis highlights recurring intraday price behavior and identifies peak and off-peak market conditions.

Key Observations

• TH_SP15 Hub prices typically peak around 6:00 a.m. and 7:00 p.m., while the lowest average prices occur around 9:00 a.m.
• The ubiquitous duck shape during sunlight hours is caused by expanding solar generation.
• The forecasting model must accurately capture these intraday price dynamics to produce reliable forecasts.

The analysis confirms that electricity prices exhibit strong hourly dependence. Consequently, the built-in Hour-of-the-Day predictor is incorporated as a core feature of the AI forecasting model.

3.2 Day-of-Week Price Pattern Analysis

To further refine the feature engineering process, the analysis is extended to evaluate hourly price behavior across different days of the week. The following plot presents the average hourly TH_SP15 Hub LMP for each weekday over the six-month training period.

Key Observations

The analysis reveals distinct weekly market characteristics:

• Saturday and Sunday exhibit lower average prices than weekdays and display noticeably different intraday price profiles.
• Monday consistently records higher average prices than the remaining weekdays, while maintaining a similar intraday pattern.

These observations demonstrate that weekly market behavior has a measurable impact on price formation and should be incorporated into the forecasting model.

4.Feature and Correlation Analysis

Selecting the most informative predictor variables is a critical step in developing an accurate AI forecasting model. Feature selection begins by evaluating the statistical relationship between potential external variables and the target variable, in this case, the TH_SP15 Hub LMP.

We make extensive use of QR Feature Engineering and Data Analysis toolbox. This cuts down data analysis time to minutes. An effective predictor should exhibit a strong correlation with electricity prices while providing meaningful insight into the market mechanisms responsible for price movements. Identifying these relationships enables the AI model to anticipate both normal market behavior and exceptional price events.

The six-month CAISO Day-Ahead dataset used in this case study presents several forecasting challenges, including fuel-driven price spikes and periods of negative electricity prices. The QuantRisk data science team addresses these challenges through systematic correlation analysis and advanced feature engineering, enabling the AI model to deliver robust and accurate forecasts.

Capturing Fuel Price Effects

The historical TH_SP15 price data exhibits several periods of exceptionally high prices, most notably during the natural gas market disruptions observed at the end of July 2026. These price spikes cannot be explained by electricity demand or weather conditions alone.

To investigate these events, our data science team performed correlation analysis across a broad range of market variables. The analysis identified a strong positive relationship between regional natural gas prices and TH_SP15 electricity prices.

As illustrated in the accompanying figures, increases in natural gas prices closely align with extreme electricity price events, confirming that fuel costs are one of the primary drivers of price volatility in the CAISO market.

Capturing Renewable Generation Effects

The historical price data also reveals multiple periods of negative electricity prices, particularly during spring and early summer.

Further analysis demonstrates that these events strongly coincide with periods of exceptionally high solar generation. During these periods, abundant renewable production combined with relatively low electricity demand creates oversupply conditions, resulting in negative wholesale electricity prices.

Zoomed over an interval of about 10 days shows clear negative correlation within the days:

Correlation analysis confirms that solar generation is one of the most influential drivers of these market conditions and should therefore be explicitly represented within the forecasting model.

4.1 Feature Engineering Decisions

Based on these findings, the following enhancements are incorporated into the AI forecasting model:

• Based on this analysis, regional natural gas prices are incorporated into the AI forecasting model as external predictors, enabling the model to anticipate fuel-driven price spikes and improve forecasting accuracy during periods of market stress.
• Solar Generation Forecasts are included as external predictor variables to improve the model’s ability to anticipate renewable-driven price events and negative pricing periods.
• Outlier Treatment, a built-in capability of QR AMPS is enabled to identify and appropriately process extreme price observations during model training. This improves model robustness while preserving genuine market signals associated with highly volatile market conditions.

By combining market expertise with data-driven feature engineering, QR AMPS successfully captures both recurring market patterns and exceptional events, resulting in highly accurate and resilient day-ahead electricity price forecasts for the CAISO market.

5.The AI Forecast Model

Following the data exploration and feature engineering stages, the AI forecasting model is configured and deployed using QR AMPS.

QR AMPS is a cloud-native, no-code AI platform that enables users to build, train, optimize, and deploy forecasting models through an intuitive drag-and-drop interface. The platform automates the complete machine learning workflow – from data preprocessing and feature engineering to model optimization, execution, and forecast publishing.

5.1. Data Preprocessing

High-quality input data is essential for producing reliable forecasts. Before model training, the following preprocessing steps are configured within QR AMPS.

Gap Filling : Missing observations are automatically detected and reconstructed using configurable interpolation methods. For this case study, Linear Interpolation is selected to preserve the continuity of the hourly price series.

Outlier Detection: Extreme price observations can significantly influence model performance. QR AMPS identifies statistical outliers using a 2.6-standard-deviation (2.6σ) detection method and replaces them with locally representative values, improving model stability while preserving underlying market behavior.

Calendar Processing: Calendar intelligence is incorporated to account for holidays and non-working days. During training and forecasting, holiday profiles are automatically mapped to the most appropriate historical demand patterns, while working days, weekends, and holidays are labeled as additional inputs to the AI model.

5.2. Feature Engineering

Based on the analysis presented in the previous sections, the following predictors are configured within QR AMPS.

These predictors enable the model to learn historical price dependencies while incorporating forward-looking market information that influences day-ahead electricity prices.

5.3. Automated Forecast Execution

The objective of this model is to generate hourly day-ahead forecasts for the TH_SP15 Hub LMP.

QR AMPS schedules the entire forecasting workflow as a fully automated daily process. The complete AI pipeline – including data acquisition, preprocessing, feature engineering, model execution, and forecast publication – is executed automatically each day at 6:00 a.m. California time, ensuring forecasts are available well before the CAISO Day-Ahead Market deadline.

5.4. AI Model Configuration

Model Splitter

The Model Splitter capability enables QR AMPS to automatically partition historical data into distinct behavioral groups and train specialized forecasting models for each market profile.

Based on the exploratory analysis, three independent forecasting models are created to capture the different characteristics of weekday and weekend price behavior:

• Monday – Friday
• Saturday
• Sunday

This segmented modeling approach allows each model to specialize in a specific market profile, improving forecast accuracy compared with a single generalized model.

Model Optimization

QR AMPS supports multiple state-of-the-art machine learning algorithms. For this case study, three high-performance gradient boosting models were evaluated:

• XGBoost
• LightGBM
• NGBoost

All three models achieved comparable forecasting performance after optimization.

Each algorithm contains numerous configurable hyperparameters that control model complexity and learning behavior. For example, an XGBoost model includes parameters such as the number of estimators, maximum tree depth, learning objective, subsampling ratio, regularization coefficients, and evaluation metrics.

Identifying the optimal combination of hyperparameters through manual experimentation is computationally impractical. QR AMPS addresses this challenge through its integrated AutoML optimization engine, which automatically evaluates thousands of model configurations and identifies the optimal hyperparameter set.

This automated optimization process minimizes the risks of overfitting and underfitting while maximizing forecasting performance.

5.5. Model Training and Execution

After optimization, the selected forecasting model is trained using six months of hourly TH_SP15 Hub LMP data together with the engineered predictor variables.

Once deployed, the complete daily forecasting workflow – including data preparation, model execution, and generation of the hourly day-ahead price forecast – typically completes in approximately 3 minutes, enabling rapid and reliable operational forecasting.

6.Forecast Results and Accuracy Analysis

6.1 Probabilistic Price Forecasting

Unlike traditional forecasting systems that generate a single point forecast, QR AMPS produces a complete probabilistic forecast by estimating the full distribution of future electricity prices.

For every hourly forecast, the platform can generate 100 forecast percentiles (P1–P100), enabling market participants to quantify forecast uncertainty and evaluate the probability of different price outcomes.

These probabilistic forecasts support a wide range of trading and operational decisions, including:

• Risk-aware bidding strategies
• Battery and energy storage optimization
• Renewable generation scheduling
• Portfolio risk management
• Value-at-Risk (VaR) analysis
• Scenario and stress testing

By combining highly accurate point forecasts with a complete probabilistic distribution, QR AMPS enables traders and asset managers to optimize expected returns while effectively managing market risk.

6.2. Forecast Dashboard

QR AMPS provides an interactive visualization dashboard that enables users to monitor forecast performance and evaluate model accuracy in real time. The dashboard consolidates key forecasting metrics into a single view, allowing traders and analysts to compare AI predictions with actual market outcomes.

The dashboard includes:

• Day-Ahead TH_SP15 Hub LMP Forecast, generated and published daily at 6:00 a.m. California time. Along with selected probabilistic forecast percentiles (5th,10th, 25th, 50th. 75th, 90th and 95th).
• Actual TH_SP15 Hub LMP, published by CAISO after the Day-Ahead Market closes, enabling direct validation of forecast performance.
• Forecast Accuracy Metrics, including Mean Absolute Percentage Error (MAPE) and Mean Absolute Error (MAE), Rood Mean Square Error (RMSE) and Forecast Direction Win Rate presented pointwise through table, and as average in gauges for continuous model monitoring.

This integrated dashboard provides complete transparency into forecast quality while enabling continuous performance evaluation and model validation.

6.3. Forecast Accuracy

The AI forecasting model achieved a Mean Absolute Percentage Error (MAPE) of 8.5% for the CAISO TH_SP15 Hub LMP.

The same forecasting methodology has been successfully applied to multiple PJM trading hubs, and across other energy markets – including PJM, MISOconsistently delivering comparable levels of forecasting accuracy with model configurations tailored to the characteristics of each market.

Forecast accuracy varies across electricity markets due to differences in their underlying market fundamentals. Key factors include:

• Market price dynamics: The drivers of day-ahead electricity prices differ significantly between regional markets. In CAISO, electricity prices exhibit a stronger dependency on natural gas prices and renewable generation than in many other North American markets.
• Data availability and predictive strength: Each ISO / RTO publishes a different set of operational and market variables. The availability and predictive value of demand forecasts, fuel prices, renewable generation forecasts, outage information, and transmission conditions directly influence forecasting performance.

By combining market-specific feature engineering with automated AI model optimization, QR AMPS consistently delivers highly accurate electricity price forecasts while adapting to the unique characteristics of each wholesale energy market.

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

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.