TL;DR
This paper models generator bidding behavior in the NYISO day-ahead electricity market using a novel three-level econometric approach. It reveals that generators exhibit significant state dependence and unobserved heterogeneity in their bidding strategies, with larger generators tending to bid lower and smaller generators engaging in more speculative pricing. The findings highlight how demand conditions critically influence bidding and market prices.
What Problem Does This Paper Address?
The deregulation of electricity markets has introduced complex bidding behaviors among generators. Understanding these strategies is crucial for market efficiency, price stability, and predicting market outcomes. Specifically, this paper investigates how generators in the New York Independent System Operator (NYISO) Day-Ahead (DA) market decide how much capacity to offer and at what price. The challenge lies in disentangling the various factors influencing these decisions, including individual generator characteristics, past bidding behavior, and market conditions like demand. Previous research has explored strategic bidding, but this paper aims to provide a more comprehensive analysis by modeling bidding at the generator, group, and aggregate supply levels.
Methods
The authors employ a sophisticated econometric framework to analyze generator bidding behavior at three distinct levels:
- Generator Level: A two-stage sample selection model, extending Heckman's approach to accommodate more than two choices, is used. The first stage models a generator's choice of bidding group (based on their maximum offer price) using a dynamic ordered probit model that accounts for both true state dependence (past choices influencing current choices) and unobserved heterogeneity (individual-specific, time-invariant characteristics). The second stage estimates the offered capacity conditional on the chosen group, controlling for selection bias.
- Group Level: For each of the five identified generator groups, an offer curve is estimated using a logistic functional form. This form is chosen to capture the observed characteristics of offer curves, such as price bounds, monotonic increases, and abrupt price jumps.
- Total Supply Level: The aggregate supply curve for the market is simulated by summing the estimated offer curves of all bidding groups. This simulated supply curve is then used to forecast market clearing prices.
Datasets
The study utilizes the NYISO Hourly Bid Data for the summer of 2001 (June 1 to September 30, 2001). This dataset comprises 39,650 observations from 325 generators, providing information on their offered capacity and bidding segments.
Key Insights and Ideas
- State Dependence and Heterogeneity Matter: Generators' decisions to join specific bidding groups are strongly influenced by their past choices (state dependence) and persistent, unobserved individual characteristics (heterogeneity). This suggests that generators don't make decisions in isolation each day.
- Size Dictates Strategy: Larger generators are more likely to bid into lower-priced groups, while smaller generators tend to be the "price speculators," offering capacity at much higher prices. This aligns with the intuition that larger, more established players might aim for stable, predictable revenue, while smaller ones might seek to exploit price volatility.
- Strategic Withholding is Likely: Although officially forbidden, the analysis suggests that generators in higher-priced groups may strategically withhold capacity to influence market prices, particularly when demand is high.
- Demand is a Key Driver: The study demonstrates that changes in demand levels significantly impact generator bidding behavior and, consequently, market prices. Higher demand generally leads to higher prices, but the magnitude of this effect is nuanced.
Key Results
- Predictive Accuracy: The dynamic models for predicting generator grouping choices achieve high accuracy, with approximately 94% and 91% of choices correctly predicted by the PA dynamic and dynamic RE models, respectively.
- Offer Curve Simulation: The logistic functional form effectively captures the shape of group offer curves, with R² values consistently above 0.85.
- Market Price Replication: The simulated aggregate supply curve, derived from the model, effectively replicates most of the volatility observed in actual NYISO DA market prices during the summer of 2001.
- Demand Sensitivity: Simulations show that a 5% increase in load can lead to market price increases of 13-20%, while a 5% decrease leads to price drops of 11-14%. The price increase effect is more pronounced than the decrease effect.
Limitations and Future Work
- Data Masking: The anonymous nature of the bidding data prevents the identification of specific plants or holding companies, limiting the analysis to market-level dynamics rather than firm-specific strategies.
- Outlier Price Spike: The analysis excludes a significant price spike on August 9, 2001, as an outlier, which might mask some important dynamics.
- Future Work: The authors suggest exploring the impact of different demand assumptions (e.g., price-responsive load) on market prices, which they begin to do in their simulations. Further research could delve deeper into the specific strategies employed by different generator types and the implications for market design.
Summary updated