Trading Electrons: Predicting DART Spread Spikes in ISO Electricity Markets

We study the problem of forecasting and optimally trading day-ahead versus real-time (DART) price spreads in U.S. wholesale electricity markets. Building on the framework of Galarneau-Vincent et al., we extend spike prediction from a single zone to a multi-zone setting and treat both positive and negative DART spikes within a unified statistical model. To translate directional signals into economically meaningful positions, we develop a structural and market-consistent price impact model based on day-ahead bid stacks. This yields closed-form expressions for the optimal vector of zonal INC/DEC quantities, capturing asymmetric buy/sell impacts and cross-zone congestion effects. When applied to NYISO, the resulting impact-aware strategy significantly improves the risk-return profile relative to unit-size trading and highlights substantial heterogeneity across markets and seasons.

💡 Research Summary

The paper tackles two intertwined challenges faced by virtual traders in U.S. wholesale electricity markets: (i) forecasting when and where extreme day‑ahead versus real‑time (DART) price spreads will occur, and (ii) converting those forecasts into profit‑maximizing trade sizes while accounting for market impact. Building on the single‑zone framework of Galarneau‑Vincent et al. (2023), the authors extend the methodology in four major directions.

First, they develop a multi‑zone, two‑sided spike‑prediction model for NYISO (11 zones), ISO‑NE (8 zones) and ERCOT (4 zones). Using only information available before each market’s gate‑closure (load forecasts, lagged DART values at 24 h and 48 h, forecast errors, calendar dummies, and seasonal indicators), they construct a 50‑dimensional feature vector for each NYISO observation (the dimension varies across markets). Binary labels for positive and negative spikes are defined by market‑specific thresholds γ pos and γ neg, calibrated to isolate economically relevant extremes. Logistic regression is chosen after extensive benchmarking against random forests, gradient‑boosted trees, and neural networks; the latter overfit the rare‑event data, while logistic regression delivers stable, interpretable probability estimates that translate well into trading signals.

Second, the authors introduce a structural price‑impact model derived from day‑ahead bid‑stack data. They estimate system‑wide impact coefficients (β sys) and zone‑specific congestion sensitivities (β zone) that map a vector of virtual load shifts q (positive for INC, negative for DEC) to expected price perturbations ΔP ≈ β sys·∑q + β zone·q_z. The model captures asymmetric buy/sell impacts and cross‑zone interactions, allowing the trader to anticipate how a large virtual position in one zone will affect both the system price and neighboring zones.

Third, they formulate a quadratic optimization problem that maximizes expected profit minus impact cost: maximize  ∑_z E