Closed-Loop Deep Learning Improves E-Commerce Pricing and Forecasting

Published in Discover Computing in October 2026, the study addresses a common limitation in existing operational research, where forecasting and pricing are typically treated as separate, isolated tasks. By coupling these two processes, the framework aims to reduce the accumulation of forecast bias and the lag in pricing strategies that often occur when the two systems do not communicate in real time.

The core of the proposed system involves two distinct but interconnected modules. The demand forecasting module is built on a Transformer architecture, which outputs both a point estimate of future demand and a price gradient. Simultaneously, a dynamic pricing module utilizes Proximal Policy Optimization, a deep reinforcement learning algorithm, to optimize continuous price adjustments. The critical innovation lies in the feedback mechanism: actual demand data generated by pricing decisions are continuously fed back to update the forecasting model, creating a cycle of mutual promotion between prediction and action.

To evaluate the framework, the study utilized two public e-commerce datasets: the UCI Online Retail Dataset and the M5 dataset. The UCI data contains approximately 540,000 transaction records from a UK online retailer between late 2010 and late 2011. The M5 dataset, published by Wal-Mart, covers 3,049 products from 10 stores in three US states from early 2011 to mid-2016. Both datasets were aggregated into weekly sequences, with outliers removed and demand values normalized using item-level z-scores to ensure comparability across different product scales.

In the demand forecasting task, the Transformer-based module demonstrated superior performance compared to several baseline methods, including ARIMA, LSTM, standard Transformer, Temporal Fusion Transformer, and Informer. On the UCI dataset, the proposed method achieved a Root Mean Square Error of 15.7, which was 16.5% lower than the best-performing baseline, Informer. On the M5 dataset, the RMSE was 22.3, representing a 13.9% improvement over Informer. These results indicate that explicitly modeling the interaction between volume and price within the forecasting process significantly enhances accuracy, particularly for high-price ranges where demand elasticity is higher.

The dynamic pricing experiments were conducted on the M5 dataset within a normalized simulator environment. The collaborative framework achieved a cumulative revenue score of 883, which was 5.0% higher than the separated scheme where the forecasting module was fixed after pre-training. This performance exceeded other reinforcement learning baselines, such as DDPG and SAC, as well as heuristic and fixed pricing strategies. The study notes that these revenue figures are normalized simulator metrics intended for relative comparison only, rather than absolute monetary values, as the simulator approximates demand responses from historical data without capturing all real-world complexities like stock-outs or competitive reactions.

Ablation experiments were performed to quantify the specific contributions of the framework’s design elements. Removing the price gradient feedback from the pricing module resulted in a 2.8% decrease in revenue, highlighting the value of providing the pricing agent with direct information about demand sensitivity. Removing the continuous update mechanism for the forecasting module led to a 4.8% revenue drop, confirming that the ability to adapt to policy-induced demand drift is crucial. When both mechanisms were removed, the system degraded to the separated scheme, with the minor difference in revenue between the two degraded configurations being statistically insignificant, suggesting that the price gradient is only effective when the underlying forecasting model is continuously updated.

The study also examined the impact of the framework across different commodity categories based on price elasticity. The revenue gains from the collaborative approach were positively correlated with price elasticity. For low-elasticity goods, such as basic food items, the gain was only 3.2%, as demand is less sensitive to price changes. In contrast, for high-elasticity goods, such as electronic products, the revenue increased by 20.5%. This suggests that the closed-loop feedback mechanism is particularly beneficial for products where small price adjustments can lead to significant shifts in consumer demand.

Sensitivity analysis focused on the update frequency of the forecasting module, denoted by the parameter K. The experiments showed that an update frequency of K=10 offered the best balance between performance and computational cost. While more frequent updates (K=2) yielded slightly higher revenue, the marginal gain was small compared to the increased computational burden. Conversely, less frequent updates (K=50 or K=100) led to significant performance drops, as the forecasting model failed to respond quickly enough to changes in the demand distribution caused by evolving pricing strategies.

The researchers acknowledge several limitations of the evaluation setup. The simulation environment relies on a non-parametric demand surface estimated from historical data, which may not fully capture the complexities of real-world pricing responses, such as competitive reactions or inventory spillover effects. Therefore, the reported revenue improvements should be interpreted as indicative relative gains within a controlled benchmark rather than direct predictions of real-world performance. The authors suggest that future work should include online A/B testing on live e-commerce platforms to validate the framework under true counterfactual conditions.

Despite these limitations, the study provides a feasible joint optimization path for e-commerce real-time pricing decisions. By breaking the traditional separation bottleneck through a closed-loop feedback mechanism, the framework effectively alleviates the impact of demand drift. The findings suggest that integrating forecasting and pricing into a unified system can lead to more robust and adaptive pricing strategies, particularly for high-elasticity goods. This approach offers a technical foundation for developing more intelligent pricing systems that can respond dynamically to market changes.

The technical implementation of the framework involved a Transformer encoder with four layers, eight attention heads, and an embedding dimension of 128. The pricing module used a three-layer fully connected network for both the actor and critic components of the PPO algorithm. The system was trained using a combination of offline pre-training on historical data and online fine-tuning during joint simulation. The alternating training schedule, where the forecasting module is updated every ten pricing decision cycles, ensures that both modules remain synchronized and responsive to the evolving market state.

The study contributes to the field of operations management by demonstrating the practical benefits of deep learning in dynamic pricing and demand forecasting. It highlights the importance of considering the bidirectional coupling between pricing and demand, rather than treating them as independent variables. The results support the use of advanced deep learning architectures, such as Transformers and reinforcement learning, to handle the non-stationary and complex nature of e-commerce demand. This work opens up new avenues for research in joint optimization and real-time decision-making in retail environments.

To prepare the datasets for evaluation, the researchers applied standardized cleaning procedures to address observational retail distortions. For the United Kingdom retail records, transactions marked with cancellation codes were converted into negative values and subsequently reset to zero, while extreme outliers across both datasets were filtered using an interquartile range threshold of three times the interquartile range. Weeks with zero demand were omitted during percentage error computations on the Walmart data to prevent division errors. To establish a controlled simulation testbed without assuming a rigid parametric formula, the authors estimated an empirical demand surface directly from observed price-demand pairs using non-parametric kernel regression. Unobserved counterfactual prices were interpolated using locally weighted regression with a Gaussian kernel and Silverman’s rule for bandwidth selection, allowing relative strategy evaluations without modeling unobserved inventory spillover.

Rather than directly modifying the reward function or constraining pricing boundaries, the automatic differentiation price gradient calculated by the forecasting module is concatenated directly onto the state vector as an informative feature. This architectural design provides the pricing agent with local demand sensitivity cues without altering the underlying revenue reward. In comparative forecasting benchmarks against price-augmented baselines, the full framework consistently retained an advantage. It surpassed price-augmented Informer by 7.6 percent in root mean square error on the British retail dataset and by 6.3 percent on the Walmart data, with both improvements achieving statistical significance at the five percent level. During training on an Nvidia A100 graphics processor, the proposed architecture converged to a loss value of approximately 0.08 within 40 epochs, outperforming competing architectures that required 60 epochs to reach higher final loss values.

The investigation also tracked cumulative regret against an omniscient hindsight benchmark constructed by assessing one hundred discrete price points per cycle. The collaborative system maintained the lowest cumulative regret, leveling off after initial exploration while fixed pricing accumulated steep regret exceeding baseline strategies. Rigorous statistical tests across five matched random seeds further confirmed the interaction between modules. Comparing the complete architecture against an ablation lacking gradient feedback showed a significant mean difference of 25.0 revenue units, whereas freezing the forecasting model produced a 42.0 unit shortfall. Meanwhile, the negligible one-unit difference between a frozen-model configuration and a completely uncoupled setup proved statistically indistinguishable with a p-value of 0.56, demonstrating that static price sensitivity gradients offer no empirical benefit when detached from continuous model adaptation.

Subject of Research: Deep learning framework for integrated e-commerce dynamic pricing and demand forecasting

Article Title: Research on the application of deep learning in E-commerce dynamic pricing and demand forecasting

Article References: Song, C. (2026). Research on the application of deep learning in E-commerce dynamic pricing and demand forecasting. Discover Computing, 29(1), Article 776. https://doi.org/10.1007/s10791-026-10745-2

Image Credits: AI Generated

DOI: 10.1007/s10791-026-10745-2

Keywords: e-commerce, dynamic pricing, demand forecasting, deep learning, Transformer, Proximal Policy Optimization, reinforcement learning, closed-loop system, UCI Online Retail, M5 dataset, price elasticity, operations management, retail analytics, forecast bias, non-parametric simulation