Abstract
In competitive e-commerce environments, sellers face the "curse of dimensionality" when attempting to optimize prices against numerous competitors. Traditional models often fail due to the massive state space and the difficulty of predicting competitor reactions. This paper proposes a computationally efficient heuristic that prioritizes high-frequency price adjustments over complex long-term market forecasting. By decomposing the optimization problem and utilizing a data-driven demand model, the strategy demonstrates a profit increase of over 20% in real-world Amazon sales experiments.
Introduction
Modern e-commerce platforms like Amazon require sellers to manage thousands of products in real-time. The primary challenge is the "curse of dimensionality": if a pricing model attempts to account for every competitor’s price, rating, and shipping time, the number of possible market states becomes mathematically intractable. Furthermore, most models assume "perfect anticipation"—the ability to predict competitor reactions—which is rarely possible in stochastic, real-world markets. This research seeks a "good enough" pricing strategy that is scalable and robust against unknown competitor behaviors.
Methodology
Demand Estimation via Logistic Regression
The authors utilize a logit model to estimate the probability of a sale based on current market conditions. Rather than tracking every individual competitor, the model focuses on key relative features that drive consumer behavior. These include the seller's price rank, the price gap relative to the cheapest competitor, and the total number of competitors.
The "Sticky Price" Heuristic and HJB Equations
To overcome the dimensionality curse, the authors introduce a stochastic dynamic programming model using a decomposition approach. Instead of solving for global market evolution, they focus on the current state and compute the optimal price for a single period using Hamilton-Jacobi-Bellman (HJB) equations.
A core component is the "Sticky Price" Heuristic: the model assumes the market is temporarily stable to calculate an optimal price, then compensates for this simplification by re-calculating at very high frequencies. This allows the seller to react to market changes as they happen rather than trying to predict them in advance. Additionally, the model incorporates adjustment costs ($z$) to prevent excessive "price flickering" that might confuse customers.
Key Insights and Results
The Power of Rank and Frequency
A critical finding is that adjustment frequency can be more important than price anticipation. A seller who reacts instantly to market shifts can outperform one who attempts to perfectly forecast the future. The demand model highlights that "price rank" is a dominant feature; moving from the second-cheapest to the cheapest position often results in a non-linear spike in sales probability.
As shown in the profit function visualization, the discrete jumps in competitor prices create multiple local maxima. This justifies aggressive strategies like "one-cent undercutting" to secure a better price rank and maximize expected profit.
Performance Evaluation
In simulations, the heuristic achieved 93% to 99% of the profit earned by "perfectly informed" strategies. More importantly, in a live experiment on Amazon Germany involving 3,300 books, the data-driven strategy outperformed an experienced professional seller’s rule-based system by more than 20% in profit. The system's efficiency allows for 2–10 updates per product per day, making it highly scalable for large inventories.
Conclusion
By shifting the focus from perfect market anticipation to high-frequency reactivity, sellers can effectively bypass the curse of dimensionality. The proposed heuristic provides a practical, data-driven framework for dynamic pricing that is both computationally efficient and significantly more profitable than traditional rule-based systems. Future work may incorporate Hidden Markov Models (HMMs) to estimate unobservable competitor variables, such as inventory levels, to further refine these strategies.
Summary updated