Capacity-Withholding Screens

• Capacity-withholding screens are statistical and algorithmic tools designed to detect both economic and physical withholding in electricity markets.
• They use a combination of statistical, structural, and behavioral metrics, including ensemble machine learning, to flag anomalous bidding patterns with high accuracy.
• Applications in day-ahead and balancing markets help regulators distinguish between legitimate hedging and anti-competitive practices to ensure market efficiency.

Capacity-withholding screens are tools, typically statistical or algorithmic, designed to detect deliberate reductions in available generation or storage capacity offered into electricity markets. Capacity withholding can arise via economic withholding—bidding resources at prices so high they are unlikely to be dispatched—or physical withholding, where owners simply do not offer available capacity. These behaviors serve as mechanisms for manipulating prices or profits, undermining market efficiency and competition. Capacity-withholding screens are essential for market operators, regulators, and competition authorities to identify, monitor, and deter such manipulative conduct, especially in markets with high penetration of storage and increasing price volatility.

1. Forms and Economic Rationale of Capacity Withholding

Capacity withholding can be instantiated as either economic or physical withholding. Economic capacity withholding involves submitting bids at above-competitive prices, effectively removing capacity from consideration without withdrawing it physically. Physical withholding is the direct non-offering of capacity.

For energy storage, the economic rationale for such withholding extends beyond classic market power exploitation. Modern electricity markets require storage bidders to incorporate forecasts of future prices and their own opportunity costs. This leads to economically rational withholding even for competitive (price-taking) actors, as they hedge against price uncertainty and maximize intertemporal profit. For example, the discharge offer function for storage is:

$O_t(p_t) = [c + (1/\eta) v_t(e_t)]^+$

where $c$ is discharge cost, $\eta$ is efficiency, and $v_t(e_t)$ is the marginal value of energy at state $e_t$ (Qin et al., 8 Mar 2024).

2. Methodological Foundations of Capacity-Withholding Screens

Capacity-withholding screens are quantitative indicators constructed from offer and dispatch data to systematically flag periods or market intervals with anomalous withholding. Foundational approaches can be divided as follows:

• Statistical screens: Metrics computed from bid and offer distributions, including spread, variance, kurtosis, and total offered/accepted quantity.
• Structural screens: Checks on the number of bids or offer blocks, and differences from historical or expected values.
• Behavioral screens: Identification of patterns consistent with strategic capacity withholding, such as persistent high bids not explained by marginal costs or opportunity values.

Recent methodological advancements employ ensemble machine learning models to combine these screens as predictors for collusion or withholding classification. For example, supervised models (Random Forest, LASSO, SVM) are trained on labeled instances of collusive versus competitive behavior using screen variables as features (Proz et al., 13 Aug 2025).

Example Capacity-Withholding Screens

Screen Type	Example Metric	Indicative Pattern
Total Offers	Number of bids submitted	Significant drop signals withholding
Offered Quantity (MW/h)	Aggregate offered capacity	Lower quantity implicates withholding
Accepted Offers/Quantity	Awarded blocks and volume in dispatch	Drop despite available capacity
Bid Distribution Statistics	Spread, skewness, kurtosis of submitted prices	High spread and skew toward upper range

3. Application in Spot and Balancing Markets

The application of capacity-withholding screens differs between market segments:

• Day-ahead (MGP) markets: Participation is optional, making the number of offers and total volume sensitive to withholding. Screens tailored for this context (e.g., total offered and accepted volume in the MGP) improve detection of complete cartels and systematic withholding (Proz et al., 13 Aug 2025).
• Balancing (MSD) markets: Participation may be mandatory, and classical screens (variance, coefficient of variation) of offer prices have been used to detect collusion and withholding.

Combined approaches, leveraging both sets of screens, yield higher detection accuracy under complete collusion. For instance, in the Campania and Brindisi cartel cases, integrating day-ahead quantity-based screens increased detection accuracy from 61% (classical screens) to as high as 95% for complete cartels (Proz et al., 13 Aug 2025).

4. Theoretical Limits and Policy Implications

The impact and bounds of economic withholding screens depend on underlying market structure and regulatory constraints:

• Unbounded Price Uncertainty: If storage operators face future prices with bounded expectations but unbounded standard deviation (e.g., Gaussian with high variance), the theoretical bid values (and hence withholding) can be unbounded. This implies no screen can fully cap rational economic bidding motivated by extreme price risk (Qin et al., 8 Mar 2024).
• Price Floor/Ceiling: Imposing explicit price bounds leads to explicit upper bounds for potential withholding, as represented by analytical expressions involving the maximum forecast price $\overline{\lambda}$ and a weighting parameter $\alpha_t$. Regulatory enforcement of such bounds provides a physical basis for screening excess withholding.

A notable policy implication is that, contrary to models assuming all withholding is anti-competitive, screening must distinguish between "beneficial" withholding (stemming from legitimate hedging against uncertainty) and exploitative withholding (market power abuse). Simulation evidence from the ISO New England system shows that, under certain stochastic assumptions, economic withholding by storage can lower system costs and align with social welfare maximization (Qin et al., 8 Mar 2024).

5. Ex-Post Analysis and Detection Protocols

Ex-post screening protocols systematically compare observed bidding and dispatch behavior against competitive benchmarks to detect potential abuse:

• Single Partial Interval Condition: For competitive price-taking storage, there should be at most one market interval where operation is at less than full capacity (outside of idling). Multiple partial intervals are indicative of potential market power exploitation (Wu et al., 2 May 2024).
• Price–Decision Relationship: Patterns where full discharging occurs only at highest prices (and full charging at lowest) fit competitive predictions. Deviations signal manipulative profitability-seeking (Wu et al., 2 May 2024).
• Statistical Testing: Non-parametric tests such as the Kolmogorov–Smirnov statistic are employed to compare bid distributions between normal and price spike days, quantifying the extent and intent of withholding (Ma et al., 23 Jan 2025).

6. Empirical Evaluations and Machine Learning Integration

Capacity-withholding screens have been validated through both empirical analysis and machine learning approaches:

• Empirical Evidence: Withholding by storage in CAISO aligns with price spikes and bid inflation, with pronounced daily periodicity and strong statistical isolation between spike and non-spike days (Ma et al., 23 Jan 2025).
• ML-Based Detection: Ensemble classifiers trained on screens achieve up to 95% (complete cartels) and 98% (incomplete cartels) classification accuracy. The incremental advantage of novel screens is highest for complete cartel detection. For incomplete collusion, classical screens suffice (Proz et al., 13 Aug 2025).

This ML-based workflow enables early warning, enhances market surveillance, and supports regulatory enforcement with quantitative evidence of manipulative intent.

7. Limitations, Open Challenges, and Future Directions

Capacity-withholding screens—statistical and algorithmic—face inherent limitations:

• Generalizability: Existing empirical studies are limited in jurisdictional scope and sample size, potentially limiting wider applicability (Proz et al., 13 Aug 2025).
• Ambiguity in Economic Withholding: Not all withholding detected by screens is anti-competitive in nature; some instances serve legitimate hedging functions in response to price uncertainty and risk sharing.
• Incomplete Collusion: New screens add little value over adapted classical screens for incomplete cartels.

Future research directions include expanding databases across more markets and periods, integrating real-time operational data, refining detection for partial collusion, and exploring alternatives to existing ensemble learning techniques to better characterize ambiguous withholding behavior (Proz et al., 13 Aug 2025).

---

In summary, capacity-withholding screens provide a rigorous framework for detecting, monitoring, and regulating both economic and physical withholding in electricity markets. Their design and interpretation require careful attention to market specifics, underlying economic rationale, theoretical risk bounds, and regulatory intent. The integration of these screens into machine learning-based systems marks a significant advance in practical market monitoring and enforcement.