MISO DAM AI Probabilistic Price Forecasting Case Study

Summary

This case study reviews a machine learning day-ahead market (DAM) energy price forecasting in MISO. All aspects of the AI forecasting pipeline are presented step by step, including, data cleansing, pre-processing, feature selection, feature engineering, probabilistic AI modeling and model hyperparameter optimization. The final AI model is executed everyday from 6 am to 9 am local time to produce a probabilistic DAM forecast. The average MAPE error is about 7.5%. The technology used to conduct this case study is QuantRisk AI platform QR AMPS (AI Market Prediction System).

Turning Market-Published Data into Actionable Day-Ahead Price Intelligence

Day-ahead market (DAM) electricity prices are among the most important signals for participants operating in competitive power markets. Generators, traders, storage operators, renewable asset owners, retailers, and other market participants rely on price forecasts to improve bidding strategies, optimize asset dispatch, manage exposure, and identify potential periods of scarcity or oversupply.

However, forecasting DAM prices is challenging. Electricity prices are not driven by a single variable. They are shaped by the interaction of demand, renewable generation, conventional generation availability, weather, fuel economics, transmission constraints, market conditions, outages and other operational factors. These relationships can also change significantly by hour and by day.

This case study used QuantRisk’s AI forecasting platform QR AMPS (AI Market Prediction System) and its latest probabilistic DAM price forecasting model for the MISO market, using six months of market-published data from December 2025 to May 2026.

The objective was not simply to build another point-forecasting model. Instead, the study focused on building an end-to-end forecasting workflow that can identify the most meaningful market drivers, generate probabilistic price forecasts, evaluate forecast accuracy, and automate the forecasting process through a cloud-based AI platform.

1. A Brief Look at the MISO Market

The Midcontinent Independent System Operator (MISO), coordinates generation and transmission resources across a broad geographic footprint. With an average 80,000 MWh load, it operates the second largest wholesale electricity market in the United States, and in the world, outside China. For market participants, the MISO Day-Ahead Market provides an important forward-looking price signal. Market participants submit offers and bids ahead of the operating day, and the market-clearing process produces hourly day-ahead prices that reflect expected system conditions.

From a forecasting perspective, this creates a particularly interesting problem. The price observed for a particular hour is the result of many interacting market conditions:

• Electricity demand and load requirements
• Wind generation
• Solar generation
• Conventional generation availability
• Supply-demand balance
• Transmission and congestion conditions
• Weather-driven demand changes
• Renewable intermittency
• Market scarcity and oversupply conditions

This means that a useful DAM forecasting system needs to go beyond historical price patterns. It needs to understand the relationship between market fundamentals and price formation.

This case study had four primary objectives:

• Understand the underlying behavior and seasonality of MISO DAM prices.
• Identify which market-published variables provide the strongest predictive information.
• Develop a probabilistic AI forecasting approach rather than relying only on a single expected price.
• Demonstrate an automated cloud workflow that can move from data ingestion to feature engineering, model development, forecasting, and performance monitoring.

2. The Data

The forecasting workflow starts with data fetched from MISO public data feed APIs. Rather than treating price as an isolated historical time series, the platform brings together multiple market time series data and transforms them into forecasting features.

2.1. Input: Market-Published Data

The input layer contains the relevant MISO market for about 6 months from December 2025 until May 2026, required to understand historical price behavior and its underlying drivers. Examples include:

• DAM price
• Electricity demand/load
• Solar generation
• Wind generation

2.2 Output: Probabilistic DAM Price Forecast

The final forecast output is not limited to a single predicted price. Instead, the QR AMPS AI forecasting system produces a probabilistic forecast, providing an expected average price prediction together with a range of potential outcomes.

3. Data Exploration: Understanding MISO DAM Price Behavior

Before developing the forecasting model, the first step was to understand the structure of the historical price data. Electricity prices are highly seasonal and exhibit strong intraday and weekly patterns. Understanding these patterns provides an important foundation for feature engineering and model development.

3.1 Raw DAM Price Data

The raw time-series data provides an initial view of how DAM prices evolved throughout the six-month study period. The visualization highlights the natural variability of electricity prices, including periods of relatively stable prices as well as sharper movements.

The key observation is that the price series is not stationary in a simple sense. Price levels and volatility vary across hours and days, which makes a purely historical-price-based approach less attractive for forecasting future market conditions. This reinforces the importance of incorporating market fundamentals into the forecasting process.

3.2 Average Hourly Price Pattern Analysis

Aggregating the historical data by hour of the day reveals a strong intraday structure. The average hourly profile shows two important peaks:

• Morning peak around 7:00 A.M.
• Evening peak around 6:00 P.M.

These patterns are consistent with changes in electricity consumption throughout the day. The morning increase reflects the rapid increase in system demand as residential activity begins. The evening peak represents another period of elevated system demand before the decline later in the night.

The presence of these repeatable patterns demonstrates why hour-of-day is an important forecasting feature of the Ai model. However, the hourly profile alone cannot explain all price movements. The magnitude of each peak can change depending on underlying demand, renewable generation, and other market conditions. This is where fundamental market features become important.

3.3 Average Day-of-Week Price Pattern Analysis

The next layer of exploration examines how the hourly profile changes across different days of the week. The analysis reveals meaningful differences between weekdays. In particular:

• Mondays show higher price levels and stronger peaks.
• Tuesdays generally follow with elevated levels.
• Wednesdays continue to show relatively strong behavior.
• Fridays show the lowest levels among the weekdays, particularly until approximately 3:00 P.M.

This demonstrates that the AI forecasting model must take into account both intraday and intra-week seasonality.

3.4 Weekend Behavior

The weekend profiles show another interesting characteristic. Saturday and Sunday do not simply represent a lower version of the weekday curve. Their behavior differs particularly during the first half of the day. The early-day patterns on Saturday and Sunday differ from the weekday profile, reflecting the different demand structure associated with weekends.

This is important for forecasting because calendar effects are not merely categorical labels. They can represent changes in the underlying market demand profile. The exploration therefore supports the use of multiple time-related features, including:

• Hour of day
• Day of week
• Weekend/weekday indicator
• Historical price behavior

4. Feature Analysis Platform: Finding the Drivers of DAM Price

Historical price patterns provide useful information, but the most valuable forecasting signal often comes from understanding why the price changes. To investigate this, the market-published variables were passed through a feature analysis workflow. The objective was to evaluate the relationship between the available market variables and MISO DAM MCE. This feature analysis process is the major task of the data science team. By natureIt it can be very time consuming and error prone when performed manually.

The QR AMPS platform has a specialized no-code Feature Extraction toolbox. This enables users to explore relationships between the target price and multiple candidate features without coding or scripting, nor requiring manual analysis of every variable. The toolbox intuitive web interface allows users to load as many potential candidates for key features, even define nonlinear functions amongst them such as ratios. Then multiple algorithms can be executed to rank the feature by importance. The process takes just a few minutes and can be iteratively repeated by adding or removing feature.

Features are ranked by importance. In practical terms one should only select the first few highest ranked features as too many features introduce noise and lower the precision of the final AI model. Also keep in mind that negatively correlated features are as valuable as positively correlated features.

4.1 Correlation with Demand

Electricity demand is one of the fundamental drivers of power prices. As demand increases, additional generation resources may need to be brought into the market, potentially increasing the marginal cost of electricity. The analysis confirms a meaningful relationship between demand and MISO MCE DAM price.

However, demand alone does not fully explain price. Two hours with similar demand can experience different prices if renewable generation or other supply-side conditions are substantially different. This led to the investigation of renewable generation.

4.2 Correlation with Solar Generation

Solar generation can significantly alter the net amount of conventional generation required to meet demand during daylight hours. Higher solar output can reduce the amount of residual demand that must be supplied by other generation resources. The feature analysis therefore evaluates the relationship between solar generation and MISO MCE DAM price.

To look more closely we zoomed on an interval of about 10 days. Here the reverse relationship between hours with high solar generation and low MCE value is crystal clear.

The relationship is not necessarily linear across all operating conditions. Solar generation interacts with the time of day, demand profile, and other generation resources, making it more useful as part of a combined market-balance feature than as an isolated predictor.

4.3 Correlation with Wind Generation

Wind generation provides another important source of variability on the supply side. When wind output is high, the amount of demand that needs to be served by the remaining generation fleet decreases. The analysis therefore evaluates the relationship between wind generation and DAM price. Again, the relationship becomes more meaningful when renewable generation is considered together with demand as a ratio.

4.4 Demand Minus Wind and Solar: A Stronger Market Driver

One of the most important findings from the QR AMPS feature analysis displayed below image is that a linear combination of features

Demand − Wind Generation − Solar Generation

provides a more useful representation of the market balance than demand alone. This feature approximates the amount of demand remaining after accounting for renewable generation.

This provides a more market-oriented representation of the conditions that conventional generation must respond to. The feature therefore combines two sides of the market simultaneously, namely demand pressure and renewable supply pressure, into demand suppression by renewables.

When demand rises while wind and solar output remain low, net demand increases, potentially placing upward pressure on DAM prices. Conversely, strong wind and solar generation can reduce net demand and potentially place downward pressure on prices.

Key finding

The QR AMPS feature analysis demonstrates an important principle for electricity price forecasting:

The best forecasting feature is not necessarily the variable with the strongest standalone correlation with price. It can be a derived feature that better represents the underlying market mechanism.

This is where AI-driven feature engineering becomes particularly valuable. Instead of simply feeding every available market variable into a model, the forecasting workflow can identify relationships that more closely reflect electricity market fundamentals.

5. End-to-End Cloud AI Forecasting Platform

Identifying useful features is only one part of the forecasting challenge. For a forecasting solution to deliver operational value, the entire workflow needs to be repeatable, scalable, and automated.

The case study therefore uses an end-to-end cloud AI platform, QR AMPS , designed to manage the forecasting lifecycle from market data ingestion through automated forecast generation. The workflow can be managed through a drag-and-drop environment, reducing the need for users to build and maintain individual data-processing scripts.

5.1. Data Preprocessing

The workflow begins with the data layer. Market-published datasets can be connected to the platform and organized within the DataHub. This creates a centralized data platform for managing the historical and incoming market data required by the forecasting process.

Electricity market data is fetched in real time and stored in the database as received. However, market operator APIs such as MISO can publish data with delays, gaps, or erroneous values, making data cleansing an essential step before AI modeling.

Gaps are AI killers. Outliers are a major source of AI forecasting error.

QR AMPS, our data engineering toolbox, provides 30+ data preprocessing methods to cleanse and prepare data before feature engineering and AI model training. These include:

• Gap detection and treatment
• Outlier detection and correction
• Data normalization
• Synthetic data augmentation, generating statistically consistent data at levels not historically observed but consistent with current observations.

This preprocessing layer ensures that AI models are trained on clean, consistent, and statistically reliable data, improving forecast robustness and accuracy.

5.2. Feature Engineering

Once cleansed data is available, the QR AMPS platform transforms raw market information into forecasting features.

This can include:

• Time-based features such as Hour, Weekday etc.
• Lagged price variables such as previous day same hour
• Historical price patterns
• Demand features
• Renewable generation features
• Net-demand features
• Other market-derived variables
• Outages

The feature-engineering layer bridges raw market data and AI model inputs. Creating, modifying, testing, and recycling features to identify the optimal feature set for a specific AI price model is traditionally a time-consuming, tedious, and error-prone process.

Automating this pipeline is therefore critical for any data science team. QR AMPS significantly simplifies feature engineering by automating the entire process and eliminating the need for Python coding, allowing data scientists to rapidly create, test, refine, and deploy feature sets for AI forecasting models.

5.3 AI Model Development and Fine-Tuning

The engineered features are then passed to the AI modeling layer, where different machine learning algorithms and libraries can be tested and backtested to identify the best-performing model for the data. This process can become highly tedious because different AI models use different hyperparameters, and the same parameters may be organized and implemented differently across machine learning libraries. Switching between models and libraries can therefore require significant coding and configuration effort.

QR AMPS eliminates this complexity through a visual, no-code modeling environment that allows data scientists to switch between AI models and libraries, configure their hyperparameters, execute backtests, and compare forecast performance within a unified interface.

For this case study, QR AMPS was instrumental in identifying the best-performing probabilistic AI model for MISO DAM price forecasting.

5.3.1 Hyperparameter Tuning

Model selection is only one step in the optimization process. The final and often most demanding task is identifying the optimal hyperparameter configuration for the selected AI model. This requires an automated process to systematically test different combinations of hyperparameters, backtest each configuration, measure forecast error such as MAPE, and iteratively refine the parameters to minimize forecasting error. QR AMPS provides an automated AutoML model fine-tuning toolbox that performs this optimization without coding. It was extensively used in this case study to fine-tune and optimize the MISO DAM probabilistic price forecasting model.

5.4 Automated Forecast Generation

The QR AMPS price forecasting platform can manage the complete workflow:

This transforms the forecasting process from a manual analytical exercise into an operational workflow.

6. Forecast Results

6.1 Forecast Accuracy Analysis

Evaluating AI model performance through quantitative error analysis and visual inspection is essential. Standard metrics such as MAPE, MAE, RMSE, and bias provide important measures of forecast accuracy, but they are not sufficient on their own. Both point and probabilistic forecasts should be visualized against actual market prices to identify trends, systematic biases, volatility, and model behavior under different market conditions.

QR AMPS automates this evaluation through its integrated price forecasting dashboard, providing error metrics, forecast distributions, and interactive visualizations of actual versus forecast prices. The screenshots presented in this case study were captured directly from this dashboard.

6.2 Probabilistic Price Forecasting

The final forecasting output provides a probabilistic view of the expected MISO DAM MCE price. Instead of producing only a single number for each forecast hour, the model provides a central forecast together with an indication of uncertainty. In the dashboard only some of the percentiles are displayed: 5th, 10th, 25th, 50th which is the main forecast, 75th, 90th, and 95th.

Moreover, as you can find in the below image, a relatively narrow prediction interval may indicate that the model has higher confidence in the expected price, while a wider interval can indicate greater uncertainty.

7. Key Findings from the MISO Case Study

The six-month MISO analysis demonstrates several important insights.

Finding 1: DAM prices have strong temporal patterns

The market exhibits clear hourly and weekly structures, including morning and evening peaks and significant differences between weekdays and weekends.

Finding 2: Demand is an important but incomplete predictor

Demand provides a meaningful signal for price, but demand alone does not capture the full supply-demand balance.

Finding 3: Renewable generation needs to be considered alongside demand

Wind and solar generation can materially change the amount of conventional generation required to serve the system.

Finding 4: Net demand provides a stronger market representation

The derived feature: Demand − Wind − Solar

provides a more useful representation of the market balance than demand alone and becomes a stronger candidate driver for DAM price forecasting.

Finding 5: Probabilistic forecasts provide more decision value than point forecasts alone

Understanding the expected price together with forecast uncertainty allows market participants to assess both opportunity and risk.

Finding 6: Automation is critical for operational forecasting

A production forecasting solution needs to manage the entire workflow—from market data ingestion to feature engineering, model execution, forecasting, and monitoring—without requiring repeated manual intervention.

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 QuantRisk automated AI electricity price forecasting solutions and the QR AMPS solution including technical description, 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.