Revolutionizing Retail with AI: Climate-Aware Demand Forecasting

By Samudrapom DamReviewed by Susha Cheriyedath, M.Sc.Sep 11 2023

In an article recently submitted to the ArXiv* server, researchers proposed a sub-neural network architecture to address the problem of seasonal climate-aware demand forecasting.

*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as definitive, used to guide development decisions, or treated as established information in the field of artificial intelligence research.

Background

Climate variability, including seasonal variability such as warmer winters or extreme events such as heatwaves, leads to significant disruption in a supply chain and impacts its resilience from inventory planning to demand management. Currently, most retailers utilize de-weatherization techniques to comprehend the demand patterns driven by weather or employ short-term weather forecasts during demand prediction due to the growing realization among them about the impact of weather in their demand forecasting.

Decision makers need reliable and accurate forecasting regarding spatial and temporal coverage to effectively perform inventory planning or demand management. This is specifically crucial while incorporating seasonal-scale forecasting into the decision-making workflow processes.

In seasonal-scale forecasting, demand forecasting evaluates seasonal climate prediction and utilizes the prediction to predict demand for several steps in the future. However, seasonal-scale climate-aware demand forecasting for time-series machine learning is challenging due to difficulties in accurate climate variability encoding for demand forecasting.

The major technical challenges in seasonal-scale climate-aware demand forecasting include the representation of mid-term to long-term seasonal climate predictions with uncertainty and the encoding of climate forecast spatiotemporal relationship with demand.

Modern deep learning (DL) techniques can partially overcome these challenges by considering climate and weather forecasts as an additional exogenous variable within the demand prediction stack. However, the consideration of such forecasts exogenous variables at the input layer can lead to unreliable and erroneous predictions for decision-making purposes owing to the high degree of uncertainty in climate prediction ensembles, which necessitated the development of effective forecasting models that consider uncertain seasonal-scale predictions and local behavior across spatial domains.

A Novel Modeling Framework

In this paper, researchers proposed a novel modeling framework that can effectively encode seasonal climate predictions to provide reliable and robust time-series forecasting for supply chain functions. The proposed framework can overcome the challenge of noisy seasonal-scale climate forecast encoding for demand prediction tasks by enabling effective learning of latent representations, such as uncertain seasonal climate prediction, through a modular neural network architecture.

The framework featured a compact uncertain seasonal climate forecast representation, which can assist in building climate resilience in supply chains by improving inventory planning and demand management. Researchers designed a modular neural network structure that accommodated various feature types, variable-length temporal window sizes based on data availability, and uncertainty related to seasonal climate forecasts. A novel method was proposed that learns the latent representations of uncertain known inputs, historical observations, and seasonal climate forecasts for seasonal-scale climate-aware forecasting.

Researchers used two types of climate encoding techniques, including the proposed latent representation learning using the sub-neural networks (LRL-SNN) technique and the temporal fusion transformer (TFT) technique, on real-world datasets to demonstrate the effectiveness of the climate-aware demand predictions. The real-world datasets include the Apparel Retail Dataset from India and the Gear Apparel Retail dataset from the United States of America (USA), two large-scale retail industry datasets, and the Favorita Grocery Retail Dataset from Ecuador, a public retail dataset.

Researchers performed these experiments to compare the results of TFT and LRL-SNN without and with climate predictions. Specifically, they evaluated the performance of their proposed LRL-SNN + Climate approach against the state-of-the-art TFT + Climate approach.

Significance of the Study

In the Favorita Grocery Retail Dataset, the incorporation of seasonal climate predictions for retail demand forecasting led to substantial improvements in model performance. 6.47% and 12.85% error reductions were observed in mean absolute percentage error (MAPE) for TFT + Climate and LRL-SNN + Climate models over the non-climate models, respectively.

Additionally, the proposed LRL-SNN + Climate approach displayed significant improvement over both error metrics, including averaged MAPE and averaged root mean squared error (RMSE) compared to TFT + Climate, which indicated that learning a temporal encoder set based on the data difficulty levels can outperform the transformer-based climate encoding architectures/TFT consistently.

In the Apparel Retail Dataset, the incorporation of climate forecasts as a latent representation part substantially improved the results for both LRL-SNN and TFT approaches, similar to the Favorita Grocery Retail Dataset. However, the TFT architecture better modeled the spatio-temporal climate variability than the LRL-SNN architecture due to the temporal attention mechanism and long short-term memory (LSTM)--based encoder-decoder architecture.

In the Gear Apparel Retail dataset, climate-aware models such as LRL-SNN + Climate and TFT + Climate displayed lower errors compared to the climate-agnostic models across stores and distribution centers. Climate-encoding using LRL-SNN led to significant improvements in error metrics compared to the TFT + Climate, and TFT approaches for store-level retail demand forecasting. However, TFT + Climate outperformed all other models for distribution centers.

Overall, the lowest MAPE value and the lowest RMSE value for both distribution centers and stores combined were realized using the LRL-SNN + Climate model and TFT + Climate model, respectively. In all datasets, adding seasonal climate predictions in LRL-SNN and TFT led to a 17% to 21% reduction in averaged MAPE and an 8% and 14% reduction in averaged RMSE.

To summarize, the findings of this study demonstrate that the latent representation of seasonal climate predictions can enhance demand forecasting, leading to improved demand management and pre-season planning for supply chain functions.

*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as definitive, used to guide development decisions, or treated as established information in the field of artificial intelligence research.