The Pitfall of “Completely Different” Markets

In DPM 101, we set up the multi-market joint assortment-pricing problem. We saw that if source markets perfectly match the target market ($\nu^{(h)} = \nu^{(0)}$), we can pool all data and learn extremely fast. If they are completely unrelated, pooling is disastrous.

But in reality, markets are rarely identical or entirely disjoint. A grocery retailer opening a store in a new city might find that customers value organic labels exactly the same way as their existing markets, but are slightly more price-sensitive.

To formalize this intuition—that markets share a lot of structure but differ in specific, meaningful ways—Chen, Chen & Zhang (2026) introduce the Utility Shift Model.

Assumption 1: Structured Preference Shift

Let the parameter vector containing both baseline preferences and price sensitivities be $\nu \in \mathbb{R}^{2d}$.

Assume there exists an index set $\mathcal{S}^* \subseteq \{1, \dots, 2d\}$ with cardinality $|\mathcal{S}^*| \le s_0$ such that, for every source market $h$:

$$ (\nu^{(0)} - \nu^{(h)})_j = 0 \quad \text{for all } j \notin \mathcal{S}^* $$

In plain English: The discrepancy between the target market and any source market is confined to a common subset of at most $s_0$ coordinates. All other $2d - s_0$ coordinates are identical across all markets.

Why this makes sense

This structure captures the operational reality that structural differences usually arise from a stable set of market-specific factors, not arbitrary idiosyncratic shifts. For instance, consumers in New York and Chicago might have identical baseline utility for a product’s technical specs, but differ materially in their baseline brand affinity or price elasticity.

The Master Knob: $s_0$

The sparsity level $s_0$ is the most important parameter in this entire framework. It quantifies the degree of cross-market similarity:

  • Small $s_0$: Markets differ along only a few dimensions. There is massive potential for information sharing.
  • Large $s_0 \approx 2d$: Heterogeneity is diffuse. The markets are structurally disjoint, and transfer learning will be ineffective.

Crucially, the algorithm does not know in advance which coordinates are shifted (the set $\mathcal{S}^*$ is unknown), nor exactly what $s_0$ is. It must discover the shifted dimensions from the data.

The Variance-Bias Tradeoff of Transfer Learning

This structured shift model perfectly sets up a fundamental variance-bias tradeoff that governs all of transfer learning:

  1. Variance Reduction (The Benefit): For the $2d - s_0$ shared coordinates, data from $H$ source markets can be safely pooled. This drastically increases the effective sample size, shrinking the statistical estimation variance.
  2. Transfer Bias (The Cost): For the $s_0$ shifted coordinates, source market data is actively misleading. If we pool it, we introduce bias. Therefore, we must rely only on target-market data to learn these specific coordinates, incurring a cost.

The overarching design principle for any algorithm in this space must be: “Shared information should be pooled to reduce variance, while market-specific deviations must be isolated to control transfer bias.”

In the next post, we will look at how the Transfer Joint Assortment-Pricing (TJAP) algorithm actually executes this principle under continuous prices and bandit feedback.