PJM DAM Price Forecast Case Study

1. Overview

The PJM Interconnection operates one of the world’s largest competitive wholesale electricity markets, coordinating the reliable delivery of power across multiple U.S. states. Electricity prices within the PJM market are calculated using Locational Marginal Prices (LMPs), which represent the cost of supplying an additional unit of electricity at specific locations, or nodes, throughout the transmission network. These prices account for energy production costs together with transmission congestion and system losses.

In addition to nodal LMPs, PJM publishes the System Marginal Price (SMP), which reflects the underlying energy component of market prices by excluding congestion and loss effects.

Market participants can trade electricity across several major PJM trading hubs, including Chicago Hub, Dominion Hub, Illinois Hub, New Jersey Hub, Eastern Hub, Western Hub, Ohio Hub, AEP GEN Hub, and AEP Dayton Hub. Trading opportunities are available in the 5-minute Real-Time Market, the hourly Real-Time Market, and the hourly Day-Ahead Market (DAM).

This case study demonstrates how QR AMPS (AI Market Prediction System), QuantRisk’s advanced AI-powered electricity forecasting platform, predicts hourly PJM Day-Ahead System Prices (DAM) using machine learning, automated feature engineering, and probabilistic forecasting techniques.

The PJM Day-Ahead Market accepts bids up to seven days before delivery and closes at approximately 12:00 p.m. on the day preceding delivery. To maximize the time available for market analysis and bidding decisions, QR AMPS automatically generates and publishes its hourly price forecasts at 8:00 a.m. market time, giving traders, utilities, and asset managers several hours to evaluate market conditions and optimize their trading strategies.

Unlike conventional forecasting systems that provide only a single expected price, QR AMPS produces probabilistic price forecasts. For every hourly trading period, the platform estimates the complete probability distribution of future prices by generating 100 forecast percentiles (P1 – P100). This enables market participants to quantify forecast uncertainty, evaluate downside and upside risk, perform scenario analysis, and develop more effective bidding and portfolio optimization strategies.

The forecasting outputs integrate seamlessly with QR TOPS (Trading Optimization & Portfolio System), where they are used to optimize bidding decisions, maximize asset revenues, and support risk-aware participation in competitive wholesale electricity markets.

The market data and forecasting results presented in this case study are sourced from QuantRisk’s public forecasting portal in this address: , where AI-based electricity price forecasts are published for PJM and numerous other North American and international power markets.

2. Market Data and Forecast Inputs

The hourly PJM DAM System Price is the core variable used to train the QR AMPS machine learning engine. Accurate electricity price forecasting depends heavily on advanced feature engineering, which helps improve predictive accuracy, model resilience, and forecast consistency under changing market conditions.

To capture the fundamental drivers of wholesale power prices, the model combines a wide range of PJM demand, renewable generation, outage, and market indicators with proprietary forecasting variables generated by QuantRisk’s AI analytics platform.

2.1 Training and Input Data

For this PJM forecasting case study, the model is trained on six months of hourly historical data, spanning the period from 1 February 2026 through 1 August 2026.

The forecasting framework incorporates several categories of input variables.

Historical Market Prices

The primary target variable is the PJM System Price published by the PJM RTO. These market prices provide the historical reference used by the AI model to learn recurring patterns, intraday behavior, and price volatility.

Electricity Demand Forecasts

Expected power consumption is represented through PJM 7 DA System Demand Forecast. This demand forecast captures anticipated load conditions across the PJM footprint and serves as a key driver of short-term electricity price movements.

Supply and Renewable Indicators

• The model also includes a set of forward-looking supply-side variables, including:
• Day-ahead wind generation forecasts
• Day-ahead solar generation forecasts
• Generation outage data published by PJM
• Outage information covers planned outages, maintenance outages, forced outages, and total unavailable capacity, providing valuable insight into generation availability and evolving market fundamentals.
• By combining these demand and supply indicators, the forecasting engine can better identify periods of tightening reserve margins, renewable variability, and potential price spikes.

2.2 Forecast Outputs

The output of QR AMPS is an hourly PJM Day-Ahead electricity price forecast for the entire following trading day and can be extended to 7 days or any on demand horizon forward.

Forecasts are automatically generated and published each day at 8:00 a.m. New York time, allowing traders, utilities, asset owners, and portfolio managers to receive actionable market intelligence well before the PJM Day-Ahead Market closes.

In addition to a traditional point forecast, QR AMPS produces a probabilistic price forecast by estimating the full distribution of possible market outcomes. For every hourly interval, the platform can generate 100 forecast percentiles (P1 – P100), enabling users to quantify forecast uncertainty, evaluate upside and downside price risk, perform scenario analysis, and support risk-aware trading and bidding decisions in wholesale electricity markets.

3. Data Discovery and Feature Engineering

Building a high-performance AI forecasting model begins with a comprehensive exploration of historical market data. Before selecting machine learning algorithms or configuring model parameters, it is essential to understand the underlying characteristics of the electricity market and identify measurable patterns that influence price formation.

To support this process, the QuantRisk data science team leverages the QR Feature Engineering & Data Analytics Toolbox, a no-code visual environment designed for advanced time-series analysis. The platform enables users to construct sophisticated data-processing workflows by connecting analytical components through an intuitive drag-and-drop interface. Behind the scenes, optimized Python libraries automatically execute the analytical pipeline, transforming raw market data into actionable insights and engineered features for AI model development.

The following analysis is based on six months of hourly PJM Day-Ahead System Price (DAM) data used to train the forecasting model.

The historical price series exhibits recurring seasonal and intraday behavior driven primarily by weather conditions, electricity demand, and changing system operating conditions. Understanding these recurring market patterns is fundamental to selecting the most informative predictor variables and designing a robust forecasting framework capable of adapting to different market regimes.

3.1 Intraday Price Behavior

A useful starting point is to examine the average hourly PJM DAM System Price across the six-month training period. Aggregating prices by hour reveals recurring daily patterns that the AI model must learn in order to produce reliable day-ahead forecasts.

Key Findings

The analysis highlights several consistent intraday characteristics:

• Wholesale electricity prices typically experience two daily peaks, occurring around 7:00 a.m. and 7:00 p.m.
• The lowest average market prices are generally observed during the early morning hours, around 3:00 a.m.
• These recurring hourly patterns indicate that electricity prices are highly dependent on the time of day.
• Rather than relying on a single model to represent every market condition, the observed behavior suggests dividing each day into four distinct operating periods:
• 03:00 – 07:00
• 07:00 – 09:00
• 09:00 – 19:00
• 19:00 – 03:00

QR AMPS automatically implements this segmentation through its Model Splitter capability, enabling multiple specialized forecasting models to be trained for different intraday market profiles. This approach allows each model to focus on a narrower range of price behavior, improving predictive performance and reducing forecasting error.

Because electricity prices exhibit strong hourly seasonality, the built-in Hour-of-the-Day predictor is also included as one of the model’s core input features.

3.2 Weekly Market Patterns

Electricity markets are influenced not only by the time of day but also by recurring weekly operating cycles. To capture these effects, the historical dataset is further analyzed by calculating the average hourly price profile for each day of the week.

This analysis reveals several statistically significant differences in market behavior.

Key Observations

The analysis identifies several distinct weekly market behaviors:

Average PJM System Prices on Saturday and Sunday are lower than weekday prices.

Thursday and Wednesday show higher average peak evening prices compared with the other weekdays.

Monday, Tuesday have higher morning peak price levels.

These observations indicate that day-of-week behavior is an important driver of PJM price dynamics and should be incorporated into the AI forecasting framework.

3.3 AI Feature Engineering Strategy

The insights obtained during data exploration directly inform the configuration of the AI forecasting model.

Based on the observed market behavior, QR AMPS applies the following feature engineering strategy:

Calendar-Based Predictors

The built-in Weekday predictor is activated to allow the machine learning model to distinguish between different weekly demand and price profiles.

Intelligent Model Segmentation

Using the Model Splitter feature, the platform automatically partitions the training dataset and builds three specialized forecasting models:

• Weekdays (Monday – Friday)
• Saturday
• Sunday

Training dedicated models for these distinct operating conditions enables the AI engine to learn the unique statistical characteristics associated with each market profile.

By combining automated feature engineering with intelligent model segmentation, QR AMPS captures both intraday and weekly market dynamics, resulting in more accurate electricity price forecasts and more reliable probabilistic predictions. In addition to forecasting the most likely market price, the platform estimates the complete probability distribution of future prices by generating 100 forecast percentiles (P1 – P100), allowing traders and asset managers to quantify forecast uncertainty and make more informed bidding decisions.

4. Feature Selection and Correlation Analysis

The performance of any AI-based electricity price forecasting model depends largely on selecting the right predictive variables. Before training the machine learning model, each potential input feature is evaluated to determine how strongly it influences the target variable, in this case, the PJM Day-Ahead System Price.

To accelerate this process, QuantRisk uses the QR Feature Engineering & Data Analytics Toolbox, a no-code environment that enables data scientists and market analysts to perform advanced statistical analysis within minutes. Through an intuitive drag-and-drop interface, users can combine mathematical operators, time-series processors, and visualization components to rapidly investigate relationships between market variables. This automated workflow significantly reduces the time required for exploratory analysis while ensuring consistent feature selection across forecasting projects.

An effective predictor should not only exhibit a high statistical correlation with electricity prices but should also represent an underlying market driver capable of explaining both routine price movements and periods of elevated volatility.

The six-month PJM Day-Ahead dataset analyzed in this case study contains a variety of complex market conditions, including sudden price spikes, changing demand patterns, and nonlinear price responses. By combining systematic correlation analysis with domain expertise, the QuantRisk data science team identifies the variables that contribute most to forecasting accuracy and incorporates them into the AI model.

4.1 Identifying Demand-Driven Market Events

One of the first observations in the historical PJM price series is the occurrence of several extreme price events, including pronounced spikes during early and mid-July. While seasonal weather conditions contributed to higher electricity prices throughout the summer, they do not fully explain these exceptional market movements.

To understand the origin of these events, the QuantRisk team performed a comprehensive correlation analysis using multiple operational and market datasets published by PJM. The analysis revealed a strong positive relationship between system demand and the largest increases in Day-Ahead electricity prices.

The accompanying charts illustrate that periods of elevated electricity demand closely coincide with the highest market prices, confirming that demand is one of the dominant drivers of price volatility within the PJM market.

These findings demonstrate the importance of incorporating demand information directly into the forecasting framework, enabling the AI model to recognize the market conditions that typically precede significant price increases.

4.2 Optimizing the Forecast Model

The insights obtained through correlation analysis are translated directly into the feature engineering strategy used by QR AMPS.

Demand Forecast Integration

The PJM 7DA Demand Forecast is incorporated as an external predictor, allowing the AI model to anticipate demand-related price spikes and better capture rapidly changing market conditions.

Intelligent Outlier Processing

Electricity markets occasionally experience extreme prices that can distort the learning process if treated as ordinary observations. To improve model stability, QR AMPS applies its built-in Outlier Treatment capability during data preprocessing.

Rather than simply removing abnormal observations, the platform intelligently identifies statistically significant outliers and processes them in a manner that preserves genuine market behavior while preventing isolated events from degrading model performance.

4.3 From Correlation to Predictive Intelligence

Feature engineering is more than selecting variables with the highest statistical correlation, it is the process of identifying the market signals that best explain future price behavior.

By combining automated feature selection, advanced statistical analysis, and electricity market expertise, QR AMPS learns both recurring market dynamics and rare high-impact events. These engineered features enable the AI forecasting engine to deliver highly accurate Day-Ahead PJM electricity price forecasts, while simultaneously producing probabilistic forecasts with 100 forecast percentiles (P1 – P100). This allows traders, utilities, and asset owners not only to estimate the most likely market price but also to quantify forecast uncertainty, evaluate price risk, and develop more resilient trading and bidding strategies.

5. The AI Forecasting Model

Once the historical data has been explored and the relevant predictive variables have been identified, the next stage is to build and operationalize the PJM Day-Ahead electricity price forecasting solution using QR AMPS.

QR AMPS is a cloud-based, no-code AI forecasting platform designed to automate the development and deployment of time-series forecasting models. Through its visual drag-and-drop environment, users can configure data processing, feature engineering, machine learning algorithms, model optimization, execution schedules, and forecast publication without writing forecasting code.

The platform brings the complete AI forecasting workflow into a single environment, from preparing raw market data through to generating and publishing production, ready electricity price forecasts.

5.1. Preparing the Market Data

Reliable electricity price predictions depend on consistent and properly prepared historical data. Before the machine learning models are trained, QR AMPS applies a series of configurable preprocessing operations to improve data quality and ensure that the input time series is suitable for forecasting.

Gap Filling

Gaps in historical time-series data can interfere with model training and reduce forecast reliability. QR AMPS provides several methods for reconstructing missing observations, including linear interpolation and historical pattern-based approaches.

For the PJM price forecasting model presented in this case study, Linear Interpolation is used to fill missing values while maintaining continuity in the hourly price series.

Outlier Treatment

Wholesale electricity markets can contain unusually high or low observations caused by exceptional operating conditions. If these observations are not appropriately processed, they can disproportionately influence the learning process.

For this model, QR AMPS uses a 3σ standard-deviation threshold to identify statistical outliers. Detected observations are replaced using a local average, helping stabilize the training dataset while retaining the overall characteristics of the PJM price series.

Calendar and Holiday Processing

Calendar effects can have a significant impact on electricity demand and consequently on wholesale power prices. QR AMPS therefore incorporates calendar intelligence into the preprocessing workflow.

The calendar processor identifies working days, weekends, and holidays and helps associate non-standard days with appropriate historical demand and price profiles. This allows the forecasting engine to distinguish normal business days from weekends and holiday periods when generating both historical training features and future forecasts.

5.2. Predictive Feature Engineering

The data analysis and correlation studies described in the preceding sections provide the basis for configuring the model’s predictive variables.

The following feature groups are implemented in QR AMPS:

These inputs allow the machine learning model to learn both historical price dependencies and forward-looking market fundamentals.

Lagged price variables capture short-term temporal relationships, while calendar predictors identify recurring intraday and weekly patterns. External demand, renewable generation, and outage forecasts provide information about expected supply-demand conditions that can influence the next day’s electricity prices.

This combination of historical, calendar, external, and derived variables provides the AI forecasting engine with a broader representation of the factors driving PJM Day-Ahead market prices.

5.3. Automated Day-Ahead Forecasting

The objective of the forecasting system is to generate an hourly PJM Day-Ahead System Price forecast for every hour of the following delivery day.

QR AMPS executes this process through an automated daily forecasting job. Once configured, the workflow runs without manual intervention and coordinates the required steps, including:

• Data acquisition
• Data preprocessing
• Feature generation
• AI model execution
• Hourly price forecasting
• Forecast publication

The automated job is scheduled according to the PJM market timetable so that the forecast is available at the required time before the Day-Ahead Market deadline. This gives traders and asset managers sufficient time to evaluate expected market conditions and incorporate the forecast into their bidding and optimization processes.

5.4. AI Model Architecture and Optimization

Model Splitter

Different segments of the electricity market can exhibit significantly different statistical behavior. Rather than forcing one generalized model to learn every pattern, QR AMPS’s Model Splitter can divide the historical dataset into distinct behavioral groups and train dedicated models for each segment.

Based on the weekly patterns identified during the data exploration stage, the PJM forecasting configuration uses three specialized models:

• Monday – Friday
• Saturday
• Sunday

This segmented approach enables each model to learn the specific demand, price, and volatility characteristics associated with its assigned group of days. It can therefore provide a more targeted representation of market behavior than a single model trained across all calendar profiles.

Automated Model Optimization

QR AMPS supports a range of modern machine learning and deep learning techniques for electricity price prediction. For this case study, three gradient-boosting algorithms were evaluated:

• XGBoost
• LightGBM
• NGBoost

Following optimization, the three approaches demonstrated comparable forecasting performance.

Each machine learning algorithm contains a large number of hyperparameters that determine how the model learns from historical data. In the case of LightGBM, for example, relevant parameters include the number of estimators, max depth, learning rate, subsampling settings, regularization coefficients, and evaluation criteria.

Testing every possible parameter combination manually would be extremely time-consuming and computationally inefficient. QR AMPS’s integrated AutoML optimization engine automates this process by systematically evaluating a large number of model configurations and identifying combinations that provide strong predictive performance.

Automated hyperparameter optimization helps reduce the risk of both overfitting and underfitting, while allowing the forecasting system to identify an effective model configuration without requiring extensive manual experimentation.

5.5. Model Training and Production Execution

After the optimal model configuration has been selected, the forecasting engine is trained using the six-month PJM hourly price history together with the engineered market predictors.

The resulting model is then deployed as part of the automated day-ahead forecasting workflow. Under the configured production setup, the complete process, from data preparation and feature generation through model execution and hourly forecast generation, typically requires approximately three minutes.

This rapid execution enables QR AMPS to produce updated PJM Day-Ahead price forecasts efficiently and consistently as part of the daily market forecasting cycle.

Probabilistic Price Forecasting

In addition to producing the expected hourly PJM price, QR AMPS supports probabilistic electricity price forecasting. Rather than limiting the output to a single point estimate, the forecasting engine can represent the range of potential future price outcomes.

For each forecasted hour, QR AMPS generates 100 forecast percentiles, from P1 through P100. This provides traders and asset owners with a richer view of forecast uncertainty and enables them to assess the likelihood of different price scenarios.

The resulting probabilistic forecasts can be used for risk-aware bidding, scenario analysis, asset optimization, portfolio management, and trading strategy development, extending the value of AI-based price forecasting beyond a conventional single-value prediction.

6. Forecast Performance and Accuracy

6.1 Probabilistic Electricity Price Forecasting

Conventional electricity price forecasting systems typically provide a single expected price for each future delivery period. QR AMPS extends this approach by producing a probabilistic forecast that represents a range of possible future electricity price outcomes.

For every hourly forecasting interval, the platform can generate 100 forecast percentiles, from P1 through P100. This provides market participants with significantly more information than a single point prediction, allowing them to assess the level of uncertainty surrounding each forecast and understand the potential range of future market prices.

Selected percentiles can be used to construct probability-based price scenarios and support a broad range of energy trading and asset management applications. By combining a point forecast with a full probabilistic representation of expected price outcomes, QR AMPS provides traders and energy asset managers with both price expectations and uncertainty information. This enables users to balance expected returns against potential market risk when making trading, bidding, and operational decisions.

6.2 Electricity Price Forecast Dashboard

The QR AMPS dashboard provides an integrated environment for monitoring forecast performance, comparing predictions with realized market prices, and tracking model accuracy over time.

Instead of requiring users to analyze multiple datasets or reporting tools separately, the dashboard brings the key forecast outputs and performance indicators together in a single visual interface.

The forecast monitoring environment includes the following components:

Day-Ahead Price Forecast

The PJM DAM System Price forecast is generated and published every day at 8:00 a.m. New York time for the following trading day.

In addition to the central forecast, the dashboard can display selected points from the probabilistic forecast distribution, including the 5th, 10th, 25th, 50th, 75th, 90th, and 95th percentiles. These quantile forecasts provide a practical view of the expected price range and associated forecast uncertainty.

Actual Market Price

Once the PJM Day-Ahead Market has closed and the corresponding market data has been published, the actual PJM System Price is displayed alongside the forecast.

This direct comparison allows traders, analysts, and data science teams to evaluate how closely the AI forecast anticipated the realized market outcome.

Forecast Performance Metrics

The platform calculates multiple statistical indicators for ongoing model evaluation, including:

• Mean Absolute Percentage Error (MAPE)
• Mean Absolute Error (MAE)
• Root Mean Square Error (RMSE)
• Forecast Direction Win Rate

The metrics are available point-by-point in tabular form, while aggregated performance indicators can also be displayed through dashboard gauges for continuous monitoring.

This combination of forecasts, realized prices, probabilistic ranges, and statistical performance indicators provides a transparent framework for evaluating the quality and reliability of the AI electricity price forecasting model.

6.3 PJM Day-Ahead Price Forecast Accuracy

For the period evaluated in this case study, the QR AMPS model achieved a MAPE of 11.3% for the PJM Day-Ahead System Price.

The same underlying AI forecasting methodology has also been applied to multiple PJM trading hubs and other wholesale electricity markets with model parameters and feature configurations adapted to the specific characteristics of each market.

Forecast accuracy can vary between power markets because electricity price formation is influenced by different combinations of demand, fuel prices, renewable generation, transmission constraints, outages, and other market fundamentals.

Two important factors are particularly relevant:

Regional Price Dynamics

The relationship between market fundamentals and electricity prices is not identical across regions. Each wholesale market has its own generation mix, transmission structure, demand profile, and price formation characteristics.

For example, CAISO Day-Ahead prices can exhibit a stronger relationship with fuel prices, while other markets may be more strongly influenced by demand conditions, renewable generation, transmission congestion, or generation availability.

Consequently, an AI forecasting model that performs well in one market cannot simply be transferred to another market without adapting its features and configuration.

Market Data Availability

The quality and breadth of data published by each independent system operator also influence forecast performance.

• Different market operators provide different combinations of:
• Electricity demand forecasts
• Renewable generation forecasts
• Fuel and commodity information
• Generation outage data
• Transmission and grid information
• Other operational market indicators

The predictive value of these variables varies from market to market. Effective feature engineering and correlation analysis are therefore essential for identifying which external time series provide the strongest signals for a particular electricity market.

6.4 Market-Specific AI Forecasting

The results demonstrate the importance of adapting an AI electricity price forecasting model to the characteristics of the market being forecast.

Rather than relying on a universal model configuration, QR AMPS combines market-specific feature engineering, automated model optimization, and specialized forecasting configurations to address the individual behavior of different wholesale power markets.

This approach allows the forecasting engine to capture recurring price patterns as well as market-specific drivers of volatility, while its probabilistic forecasting capability provides a broader view of potential price outcomes.

By generating both point forecasts and 100 forecast percentiles (P1 – P100), QR AMPS enables energy market participants to move beyond a single expected price and incorporate forecast uncertainty into trading, bidding, and asset optimization decisions.

6.5 Learn More About AI Electricity Price Forecasting

For additional information about QR AMPS, including AI-powered electricity price forecasting solutions, technical capabilities, case studies, and live interactive forecast dashboards, visit:

https://quantrisk.com/analytics-electricity-price-forecast

If you have questions about our forecasting technology or would like to discuss your electricity price forecasting requirements, please contact [email protected]. Our team will be pleased to discuss your market, forecasting objectives, and potential applications of probabilistic AI forecasting.