AI-Based SPP DAM MEC Price Forecasting

A Six-Month Case Study on Probabilistic System Price Forecasting in the SPP Day-Ahead Market

Summary

Wholesale electricity price forecasting requires capturing both the temporal structure of the market and the uncertainty associated with future price movements. This case study evaluates a forecasting workflow for the Southwest Power Pool (SPP) Day-Ahead Market (DAM), using the System Marginal Energy Component (MEC) as the target variable. The study uses approximately six months of SPP market data from January through June 2026 and evaluates forecasts for the first 10 days of July 2026. The workflow combines market-published data, contextual time-series variables, feature engineering and selection, machine-learning models, automated time-series backtesting, and probabilistic forecasting.

The resulting workflow can generate forecasts before the official Day-Ahead price is available, including a configurable 1-Day-Ahead forecast generated at 6:00 AM on the day before the operating day. Longer horizons, such as 7-Day-Ahead or other user-defined horizons, can also be generated using the same framework. In addition to a central price forecast, the system produces a distribution of possible price outcomes through 100 forecast percentiles, allowing forecast uncertainty to be represented explicitly. Model performance is evaluated using MAPE (9.9%), MAE (3.32 $/MWh), RMSE (4.54 $/MWh), and directional win rate (85.83 %).

This case study examines an AI-powered forecasting workflow powered by QR AMPS (AI Market Prediction System)

1. Six Months of SPP MEC: What Does the Market Data Tell Us?

Before applying machine-learning models, the historical SPP MEC data is analyzed to identify recurring temporal patterns and variations in price levels. The study uses approximately six months of SPP data from 2026 to train the AI model, which then forecasts MEC prices for the period from July 1 to July 10, 2026. The raw MEC time series shows substantial volatility throughout the study period.

Periods of relatively stable pricing are interspersed with sharper spiky movements and changes in daily price levels. This indicates that a simple historical-average model would not adequately represent the observed behavior. Although electricity prices contain recurring patterns, their magnitude and level vary with market conditions. The exploratory analysis therefore focused on two questions:

• Which repeatable temporal patterns exist?
• Which market variables help explain deviations from those patterns?

1.1 Intraday Pattern

As shown in the image below, the average hourly price profile reveals two distinct price peaks:

• Morning peak around 09:00
• Higher evening peak around 18:00

The evening peak is particularly pronounced, indicating that price levels vary significantly across operating hours. This reinforces the importance of including hour-of-day as an input feature in the forecasting model.

However, the average hourly curve reflects typical market behavior rather than the expected MEC for a specific operating hour. MEC values at 09:00 or 18:00 can vary substantially from day to day, depending on demand and other prevailing market conditions. These relationships are examined in the following sections.

1.2 Weekday Effects

The weekday analysis below shows that the overall hourly profiles are relatively similar, while their price levels vary across days.

The heatmap demonstrates differences in weekday-hour behavior:

Before noon:

• Monday and Wednesday generally exhibit the highest levels.

During the second half of the day:

• Tuesday and Saturday show higher levels.
• Thursday tends to have the lowest levels.

This indicates that the forecasting model needs to represent more than an hourly pattern. A forecast for 18:00 should distinguish between Monday, Tuesday, Thursday, and Saturday rather than treating these observations as equivalent. Combining hour-of-day and day-of-week information provides a more complete representation of temporal market behavior.

2. Feature Discovery and Engineering

Historical price patterns provide only part of the information required for forecasting. The next step is to identify which available market variables contain predictive information. For this study, demand was one of the primary variables investigated.

2.1 Demand vs. MEC

The simple correlation between demand and MEC is approximately 45%.

Although this is not a particularly strong linear correlation, the following more focused visualization shows that many major demand increases and decreases occur at similar times to corresponding MEC movements.

This indicates that demand contains useful temporal and directional information even though the relationship is not adequately described by a simple linear correlation. This distinction is important for electricity-price forecasting. A correlation coefficient measures one specific type of relationship and does not capture nonlinear, conditional, or time-dependent interactions that may be useful to a machine-learning model.

2.2 Average Price Across Demand Levels

The analysis therefore also examines average MEC prices across different system demand levels.

The visualization provides a complementary perspective to the correlation analysis. Rather than measuring the strength of the overall relationship between demand and price, it shows how average MEC levels vary across different ranges of system demand.

The resulting trend helps identify whether higher-demand conditions correspond to different price regimes and whether the price response changes across demand levels. This is relevant to machine-learning models because the relationship between system demand and price is not necessarily linear.

3. QR Feature Engineering & Selection Toolbox

A key feature of the QR AMPS forecasting workflow is its no-code, on-screen feature engineering and selection toolbox. It enables multiple datasets to be integrated and candidate forecasting inputs to be evaluated within a single environment. In the example below, SPP-published market data is incorporated into the workflow. Data scientists can explore, transform, and evaluate different features within minutes, without writing Python code. This accelerates feature preparation and model development while keeping the forecasting workflow accessible and reproducible.

Multiple feature-selection methods can then be used to generate a ranked list of features based on their relationships with the target variable and their selection by the applied algorithms.

For the final model, users can select relevant market variables and construct contextual features through the visual interface without manually implementing every transformation. The following image illustrates this process.

The toolbox can be used with features such as:

• Market-published demand
• Historical MEC
• Lagged values
• Rolling averages
• Rolling minimums and maximums
• Hour-of-day context
• Day-of-week information
• Weekend/weekday indicators
• Historical price statistics
• Demand transformations
• Other derived time-series variables

Rather than treating all available variables equally, the art of feature engineering is to examine relationships, create derived features, compare feature behavior, and construct a model-ready feature set within the same workflow. The process can therefore be represented as:

Raw market data → Feature transformation → Feature selection → Model-ready dataset

This also provides a consistent process for reproducing the feature-generation workflow when moving from research to production.

4. AI Model Development and Automated Backtesting

Once the feature set has been established, the next step is to evaluate models capable of learning the relationship between the available inputs and future MEC. It is highly desirable to have a forecasting workflow that supports multiple Machine Learning and Deep Learning approaches, allowing alternative algorithms to be evaluated using the same prepared dataset and features. Rather than selecting an algorithm in advance, different approaches can be tested and their configurations adjusted, including algorithm-specific hyperparameters. This changes the model-development process from simply asking:

“Which algorithm should we use?”

to evaluating:

“Which algorithm and configuration provides the best out-of-sample forecasting performance for SPP MEC?”

This case study greatly benefited from QR AMPS flexible no-code AI environment. The image below illustrates the no-code model configuration interface, where the AI models, parameters, and hyperparameters can be configured.

4.1 Importance of Backtesting

To improve performance, model selection should be supported through time-series backtesting. Historical periods can be treated as simulated forecasting exercises, with each forecast generated using only the information that would have been available at the corresponding forecast timestamp.

This is important for electricity-price forecasting because conventional random train-test splits can produce unrealistic performance estimates when information from future periods enters the training process. The backtesting workflow instead evaluates models in a manner that more closely represents the actual forecasting sequence.

Candidate AI models (algorithms and configurations) can then be compared using multiple performance measures. The final model is established using:

Market Data + Contextual Time-Series Features + Selected ML Algorithm + Optimized Configuration

The objective is therefore not simply to minimize historical fitting error, but to evaluate how well the model generalizes to previously unseen market conditions.

5. From 1DA to 7DA and Beyond

After model development and validation, the forecasting workflow can be operationalized through the QuantRisk cloud environment. The main stages of the forecasting process are connected as:

DataHub → Feature Engineering → Model → Forecast → Monitoring

Market data can be ingested into the DataHub, transformed through the feature-engineering workflow, passed to the selected model, and converted into an automated forecast. The resulting process can then be scheduled according to the required forecast timing.

5.1 1-Day-Ahead Forecast

For example:

6:00 AM → Forecast tomorrow’s 24 hourly MEC values

This produces an independent forecast before the official DAM price becomes available later in the day.

5.2 7-Day-Ahead Forecast

The same workflow can be configured for a longer forecast horizon:

Today → Forecast the next seven days

This provides additional forward information for planning and analysis.

5.3 On-Demand Forecast Horizons

The forecasting horizon does not need to be limited to 1DA or 7DA. The workflow can be configured for a user-defined horizon based on the forecasting requirement.

Different applications require different amounts of forward visibility. Short-term market decisions may use a 1DA forecast, while portfolio planning, maintenance scheduling, storage strategy, or risk analysis may require a longer horizon.

6. Probabilistic Forecasting at Scale

The forecasting output should not be limited to a single expected price. The QR AMPS system can generate 100 forecast percentiles within a few minutes. The image below shows a subset of these percentiles – P5, P10, P25, P50, P75, P90, and P95 – for visualization. The complete set of 100 percentiles is generated and stored in the database as the probabilistic forecast.

This provides a distribution of possible MEC outcomes for each forecast hour rather than only a single point estimate.

For example:

• Lower percentiles, e.g., P5, represent lower-price scenarios.
• Middle percentile P50 represents the central forecast region.
• Upper percentiles, e.g., P95, represents higher-price scenarios.

This representation is relevant in power markets because forecast uncertainty can increase substantially under unusual demand conditions, changes in generation availability, congestion, renewable generation variability, or other market events.

7. Results and Forecast Performance

Finally, putting all these together, the forecasting experiment was performed for July 1 – July 10, 2026. Results are provided through interactive reports and dashboards in both chart and tabular formats. Forecast performance is evaluated using four complementary measures:

7.1 Directional Win Rate: An Important Commercial Metric

For electricity-market forecasting, numerical price error is not necessarily the only relevant measure. A forecast can have a relatively low MAE while incorrectly predicting the direction of the next price movement. Conversely, a forecast can have a larger numerical error while still correctly identifying whether the market is moving higher or lower.

The directional win rate measures how frequently the model correctly predicts the direction of price movement according to the defined methodology. It provides a complementary measure of model performance alongside the numerical error metrics.

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 more comprehensive assessment of forecasting performance.

Conclusion

The six-month SPP study demonstrates a forecasting workflow for generating an early and probabilistic view of Day-Ahead MEC using publicly available market information.

The historical analysis identified clear intraday behavior, including prominent MEC peaks around 9 AM and 6 PM, as well as differences in price levels across weekdays. Demand showed an approximately 45% correlation with MEC, while the visual analysis demonstrated that demand movements and major price changes were often aligned. The analysis across demand levels further indicated that the relationship between demand and MEC is not necessarily linear, supporting the use of models capable of representing nonlinear relationships.

The forecasting workflow integrates feature engineering, feature selection, machine-learning and deep-learning model evaluation, hyperparameter configuration, and automated time-series backtesting. The on-screen QR AMPS feature engineering and selection toolbox allows market data and contextual time-series variables to be transformed and evaluated within the same environment.

Following model selection and validation, the forecasting workflow can be operationalized as an automated cloud process. Forecasts can be generated as early as 6:00 AM before the next DAM, or at another user-defined time. The same workflow can support 1DA, 7DA, or other forecasting horizons and can generate 100 price percentiles within minutes to represent the distribution of potential MEC outcomes.

Forecast performance was evaluated using MAPE, MAE, RMSE, and directional win rate. The first three metrics quantify numerical forecast error, while directional win rate provides an additional measure of whether the model correctly anticipated the direction of market movement.

For additional technical information, including forecasting architecture, related case studies, and interactive forecast dashboards, see the electricity price forecasting resources available from QuantRisk. For questions or comments, contact the team at [email protected].