When and How Should a Power Trader Engage in Arbitrage? Predict, then Contextually Optimize

Technical Summary: Predict-Then-Contextually Optimize for Power Trader Arbitrage

Problem Statement
Stochastic energy generators, such as wind farms, face increasing opportunities and risks in electricity markets due to widening price spreads between the day-ahead and balancing markets. Under single-price balancing schemes (common in Europe), traders are exposed to price risk regardless of the direction of their imbalance, creating potential for opportunistic arbitrage. However, existing bidding strategies often lack explainability and explicit risk management. Traders must decide not only if to deviate from their production forecast to exploit a predicted price spread ($\Delta\lambda_t = \lambda^{DA}_t - \lambda^B_t$), but also in which direction (long or short) and to what extent. Current frameworks frequently obscure the economic logic of these decisions or fail to provide a pragmatic, risk-aware mechanism to tune the trade-off between profit and tail risk (e.g., Conditional Value-at-Risk, CVaR). Furthermore, the problem is compounded for hybrid power plants (HPPs) combining renewables with assets like electrolyzers, which introduce temporal constraints and additional operational flexibility.

Methodology: A Predict-Then-Contextual-Optimize Framework
The authors propose a three-stage framework that decomposes the day-ahead bidding decision into explicit, explainable steps:

1. Stage 1: When to Engage (Probabilistic Classification)
   A probabilistic binary classifier predicts the sign of the price spread ($\Delta\lambda_t > 0$ or $< 0$) based on contextual features (e.g., weather, market prices, demand forecasts). To manage risk, the framework employs two tunable confidence thresholds ($\alpha$ and $\bar{\alpha}$) on the predicted probability.

   • If the prediction confidence is high (probability $\ge \bar{\alpha}$), the trader engages in an opportunistic long bid.
   • If the prediction confidence is low (probability $\le \alpha$), the trader engages in an opportunistic short bid.
   • If the prediction falls within the "deadband" ($\alpha < \text{prob} < \bar{\alpha}$), the trader defaults to an arbitrage-free bid (equal to the production forecast).
   • Training Detail: The classifier is trained using LightGBM. A "deadband quantile" ($\delta$) is introduced to exclude training samples with near-zero price spreads, which are high-noise and low-value. The thresholds $\alpha$ and $\bar{\alpha}$ are tuned on a validation set to maximize a value-oriented metric (expected profit or CVaR), rather than minimizing classification loss.

2. Stage 2: What Extent (Contextual Optimization)
   Conditional on the decision to engage in arbitrage (long or short), a linear decision policy determines the magnitude of the bid deviation from the forecast. Two separate policies are learned via contextual optimization: one for the long class and one for the short class.

   • The policy maps contextual features directly to a signed deviation $z_t$, such that the bid is $p^{DA}_t = \hat{P}^W_t + z_t$.
   • The optimization objective maximizes a weighted combination of mean profit and CVaR of profit, subject to physical constraints (power balance, electrolyzer operational limits, hydrogen production targets).
   • For HPPs, the model incorporates constraints for the electrolyzer, allowing it to absorb mismatches or procure grid power, thereby increasing flexibility.

3. Stage 3: Ex-Post Profit Calculation
   In the testing phase, the committed day-ahead bid is settled against realized prices. For HPPs, a secondary deterministic optimization is performed to dispatch the electrolyzer based on the balancing price forecast, maximizing the residual revenue.

Key Contributions

• Explainable Decomposition: Unlike integrated optimization models that output a single bid quantity, this framework explicitly separates the decision into "when," "direction," and "extent," making the economic logic transparent to traders.
• Risk-Aware Tuning: The framework introduces two distinct mechanisms for risk control: (1) confidence thresholds in the classification step control the frequency of opportunistic bids, and (2) the CVaR weight ($\beta$) in the optimization step controls the magnitude of the bid deviation.
• Hybrid Asset Integration: The methodology is applicable to both standalone renewable generators and hybrid plants, demonstrating how co-located assets (like electrolyzers) can enhance arbitrage value by providing operational flexibility to manage imbalances.
• Contextual Optimization: The approach leverages contextual optimization to learn linear policies that adapt to market conditions without requiring complex, non-convex solvers at inference time.

Results
The framework was evaluated using a rolling-window procedure on real data from European bidding zones DK1 (Denmark) and DE/LU (Germany/Luxembourg) for a 7.2 MW wind farm, with and without a 3.6 MW electrolyzer.

• Profit Improvement: The proposed "Classification + Policies" model increased mean profit by approximately 7% compared to a standard arbitrage-free benchmark (bidding the forecast) for the hybrid plant in DK1.
• Flexibility Value: The hybrid plant (HPP) consistently outperformed the standalone wind farm, achieving profit improvements of up to 75% in DK1 and over 100% in DE/LU, attributed to the electrolyzer's ability to absorb imbalances.
• Risk-Return Trade-off: The model achieved a higher mean profit than the "Single Policy" benchmark while significantly reducing the CVaR (tail risk). Compared to an "All-or-Nothing" strategy, the proposed model maintained comparable mean profits but drastically reduced the severity of worst-case losses.
• Distribution Drift: Performance was highly sensitive to distributional drift between training and testing windows. Profit improvements were highest when drift was low and collapsed (or turned negative) under severe drift, highlighting the limitations of static models in non-stationary markets.

Significance and Claims
The paper claims that its primary significance lies in providing a pragmatic, explainable, and risk-aware decision support tool for power traders. By decoupling the prediction of market opportunities from the optimization of bid quantities, the framework allows traders to explicitly tune their risk exposure through interpretable parameters (confidence thresholds and risk weights). The authors note that while the framework improves individual trader profits, it does not analyze the system-level impacts (e.g., on market welfare or liquidity), which they identify as a direction for future work. The study also acknowledges that non-stationarity remains a challenge, suggesting that future iterations could incorporate online learning to adapt to changing market dynamics.