Probabilistic Price Forecasting for Simultaneous Auctions

This direction is merged into another direction: stochastic price modeling. Its last version stays readable and citable here.

Version 1 of 112 papers18 conceptsfrozen 2026-10-04

In a simple auction for a single item, your strategy is straightforward: determine what the item is worth to you and bid accordingly. But in complex markets like the Trading Agent Competition (TAC), agents must bid on dozens of items simultaneously—flights, hotels, and entertainment tickets—that are only valuable if won as a complete package.

In these "interacting" auctions, traditional bidding strategies based on average expected prices fail. Instead, a lineage of research led by Peter Stone, Michael L. Littman, and Amy Greenwald demonstrates that agents must model the full probability distribution of future prices to make optimal decisions.

The Failure of Point Estimates

Most predictive models output a single number: the expected value. However, as Michael L. Littman and Robert E. Schapire argue in their work on the ATTac agent, the mean price is a dangerous simplification in volatile markets.

Consider a travel agent needing a flight and a hotel. If the expected price for each is \$200, a point-estimate model suggests a total cost of \$400. But if the hotel price is actually a bimodal distribution—either \$100 or \$500 with equal probability—the "average" of \$300 is a price that may never actually occur. If the agent bids based on the \$300 average, they risk winning a flight they can no longer afford the hotel for, or losing a profitable opportunity because they over-estimated the cost.

The Exposure Problem and Interdependence

The core challenge in simultaneous auctions is the exposure problem. As Justin A. Boyan proves in Bidding under Uncertainty: Theory and Experiments, bidding your "marginal utility" (the value an item adds to your current set) is optimal for sequential auctions but fails when auctions close at the same time.

Because the value of a hotel room is zero without a corresponding flight, an agent is "exposed" to the risk of winning only half a bundle. To navigate this, agents like RoxyBot-06 Greenwald et al. use a "predict-then-optimize" framework. Instead of asking "What is the price?", the agent asks "What is the likelihood of every possible price?" and then optimizes its bids across thousands of simulated scenarios.

Density Estimation and Monte Carlo Sampling

To turn these probability distributions into actual bids, researchers moved away from parametric models (like the Normal distribution) which often fail to capture the "fat tails" or multiple peaks of real market data.

  1. Conditional Density Estimation: Stone et al. utilized boosting-based algorithms to learn the full density of clearing prices based on current market features.
  2. Non-parametric Representations: Justin A. Boyan suggested representing price uncertainty using lookup tables of percentiles. This allows the agent to capture highly skewed distributions where the "most likely" price is far from the "average" price.
  3. Sample Average Approximation (SAA): Once the distribution is known, agents use Monte Carlo sampling to generate hundreds of "possible futures." As detailed in the RoxyBot-06 architecture, the agent finds a bid that performs well on average across all these samples, effectively hedging against market volatility.

Contextual Shading and Economic Regimes

Recent iterations of this idea have expanded the "context" used to build these distributions. In Learning to Bid in Contextual First Price Auctions, Ashwinkumar Badanidiyuru and Zhe Feng highlight that bidders must account for both systematic context (like the time of day) and the random noise of competitor behavior to "shade" their bids—finding the sweet spot between winning the auction and overpaying.

Furthermore, markets aren't static; they shift between "economic regimes." Wolfgang Ketter and Maria L. Gini demonstrated in the MinneTAC agent that by identifying whether a market is in a state of scarcity or over-supply, an agent can adjust its price density functions in real-time. This allows for risk-aware decision-making, such as identifying a "price floor" to avoid selling at a loss during a market crash Ketter et al..

By shifting the focus from predicting the future to modeling the uncertainty of the future, these agents transform bidding from a game of guessing into a rigorous exercise in stochastic optimization.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_probabilistic_price_forecasting_for_simu_f1d8ad49,
  title = {Probabilistic Price Forecasting for Simultaneous Auctions},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/probabilistic-price-forecasting-for-simultaneous-auctions/rdr_f1d8ad495d8b4e76b180778e02dd419d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_probabilistic_price_forecasting_for_simu_f1d8ad49_v1,
  title = {Probabilistic Price Forecasting for Simultaneous Auctions},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/probabilistic-price-forecasting-for-simultaneous-auctions/rdr_f1d8ad495d8b4e76b180778e02dd419d/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}