TL;DR
The researchers developed PVD-B (Pricing with Volume Discounts Bandit), a two-phase online learning algorithm that optimizes tiered pricing strategies using only raw transaction data. In a four-month real-world A/B test with an Italian e-commerce firm, the algorithm outperformed human pricing specialists by 55% in total profit and increased average basket sizes by roughly 33%.
What Problem Does This Paper Address?
In e-commerce, not all customers are created equal. Business-to-business (B2B) clients typically buy in bulk, while business-to-consumer (B2C) shoppers buy single units. Ideally, a seller would offer different prices to these groups. However, as shown in Figure 1, these classes are often indistinguishable until the moment of checkout.
The challenge is twofold:
- Price Discrimination without Identification: How do you offer lower per-unit prices to bulk buyers (volume discounts) without knowing who they are upfront?
- Data Scarcity: Learning the "perfect" price for every possible volume threshold is computationally expensive and requires massive amounts of data—something most retailers lack for individual products.
Methods
The authors propose the PVD-B algorithm, which simplifies the complex task of learning multi-tiered prices into two manageable phases (see Figure 3 for the workflow).
- Optimal Price Estimation: Instead of learning every tier at once, the algorithm first estimates a single "optimal average price" ($p^*_t$) for the product. It uses Bayesian Linear Regression (BLR) to model the demand curve, incorporating seasonality through Radial Basis Functions (RBF). To ensure the model follows business logic, they use Lognormal priors to force the demand curve to be monotonically decreasing—meaning higher prices always lead to lower or equal demand.
- Volume Discounts Learning: Once the average price is found, the algorithm distributes this price across different volume tiers. It uses a "buyback probability" ($\gamma$)—the likelihood a customer will return to buy the same batch size—to calibrate discounts. If a product has a high buyback rate (like office supplies), the algorithm is more conservative with discounts.
To handle exploration (trying new prices) vs. exploitation (using the best-known price), the authors employ Thompson Sampling, as illustrated in Figure 4.
Datasets
The algorithm was evaluated using a proprietary dataset from a major Italian e-commerce company:
- Historical Data: Two years of transaction records (date, time, product ID, units, price, cost, and customer class).
- A/B Test Set: 295 products following a "long-tail" economic model.
- Control Set: 33 products managed by human pricing specialists.
- Scale: The test involved products with a yearly turnover of approximately 300,000 Euros.
Key Insights and Ideas
- Volume as a Proxy: Since customer identity is hidden, the number of units in a basket serves as a proxy for the customer class. Volume discounts act as a self-selection mechanism for price discrimination.
- Monotonicity is a Shortcut: By mathematically forcing the demand curve to be monotonic, the algorithm converges much faster. It doesn't waste time "learning" that doubling the price might miraculously increase demand.
- Seasonality Matters: The authors found significant intra-week and yearly seasonality (see Figures 5 and 6). The algorithm updates prices every 7 days to account for the fact that people shop differently on weekends versus weekdays.
Key Results
The real-world deployment yielded impressive gains:
- Profit Increase: The PVD-B configuration achieved a 55% higher total profit than the human-managed control group.
- Basket Size: The introduction of optimized volume discounts increased the average units purchased per basket by 33%.
- Statistical Significance: Permutation tests confirmed the algorithm's superiority with at least 98% confidence.
- Operational Success: Following the test, the company expanded the algorithm's use to over 1,200 products, representing roughly 50% of their total turnover.
Limitations and Future Work
While highly effective, the current PVD-B model treats every product as an island. It does not account for cross-selling effects (e.g., how a discount on printers affects ink sales).
The authors suggest that future iterations should:
- Incorporate long-term loyalty effects more deeply into the reward function.
- Develop multi-product optimization strategies to capture cross-selling dynamics.
- Refine the buyback probability estimation to better handle highly non-stationary environments.
Summary updated