Author: Denis Avetisyan
New research shows that incorporating principles of revealed preference – even with imperfect consumer behavior – can significantly improve the accuracy of time-series forecasting models.
Fine-tuning foundation models with synthetic data derived from utility maximization using the Generalized Axiom of Revealed Preference (GARP) enhances demand prediction performance.
While modern time-series foundation models excel at forecasting without specific training, they often fail to incorporate fundamental principles of economic behavior. This limitation motivates the work ‘GARP-EFM: Improving Foundation Models with Revealed Preference Structure’, which demonstrates that fine-tuning such models with synthetic data-generated from agents maximizing utility subject to a budget constraint and adhering to the Generalized Axiom of Revealed Preference (GARP)-substantially improves demand forecasting accuracy. By leveraging GARP to create structured training data, the authors show that a fine-tuned Amazon Chronos-2 model learns robust price-quantity relationships, even outperforming zero-shot predictions on real consumer choices. Could this approach of embedding economic theory into synthetic data generation unlock further improvements in foundation model performance across diverse forecasting applications?
The Illusion of Prediction: Modeling Consumer Choice
Effective demand forecasting stands as a cornerstone of efficient resource allocation across numerous industries, yet consistently accurate predictions remain elusive due to the inherent complexities of consumer behavior. Traditional forecasting methods, often reliant on historical sales data and statistical trends, frequently falter when confronted with the nuanced and often unpredictable choices of individuals. Consumers don’t operate in a vacuum; their decisions are shaped by a multitude of factors – shifting economic conditions, evolving preferences, external influences, and even psychological biases – that are difficult to capture within simplistic models. This inability to adequately represent the underlying drivers of demand leads to inaccuracies in forecasting, resulting in overstocking, stockouts, wasted resources, and ultimately, reduced profitability. Consequently, a need exists for more sophisticated approaches that move beyond mere extrapolation and delve into the core principles governing consumer choices.
A fundamental difficulty in predicting consumer demand stems from accurately representing how individuals make choices. Conventional forecasting often treats purchasing patterns as purely statistical, overlooking the underlying assumption of rationality – that people generally strive to obtain the most satisfaction, or ‘utility’, from their limited resources. However, consumer behavior isn’t always perfectly logical; individuals are subject to biases, incomplete information, and changing preferences. Consequently, models must account for these deviations while still maintaining a core foundation in rational choice theory. Successfully incorporating this principle allows for the creation of more robust and realistic demand forecasts, moving beyond simple extrapolation of past sales to a deeper understanding of why consumers choose certain products over others, and how those choices might evolve under different conditions.
The forecasting of consumer demand benefits from a novel framework built upon established economic principles. This approach centers on the concept of revealed preference – the idea that an individual’s choices reveal their underlying preferences – and extends it through the Generalized Axiom of Revealed Preference (GARP). By computationally simulating consumer behavior adhering to GARP, researchers can generate synthetic datasets that mirror the consistency and rationality expected of real-world decision-making. This allows for the creation of controlled environments for training and rigorously testing demand forecasting models, overcoming the limitations inherent in relying solely on often-noisy and incomplete observational data. The resulting synthetic datasets provide a valuable tool for benchmarking algorithms and enhancing the accuracy of future predictions by grounding them in a solid foundation of economic theory.
The creation of a controlled environment for forecasting model development offers a significant advantage over traditional methods reliant on observational data, which is often noisy, incomplete, and subject to unforeseen external factors. By generating synthetic datasets based on principles of rational consumer choice – and specifically utilizing the Generalized Axiom of Revealed Preference – researchers can isolate the impact of specific variables and test model performance under known conditions. This allows for rigorous evaluation of forecasting accuracy and the identification of model weaknesses before deployment in real-world scenarios. Furthermore, this approach facilitates the exploration of ‘what-if’ scenarios and the assessment of model robustness to changes in consumer behavior, ultimately leading to more reliable and efficient demand predictions.
Chronos-2: Mapping the Algorithm of Demand
Chronos-2 utilizes the transformer architecture, a neural network design originally developed for natural language processing, to model and forecast time series data. This approach allows the model to capture long-range temporal dependencies, unlike recurrent neural networks which can struggle with sequences exceeding their memory capacity. The transformer’s self-attention mechanism enables each time step to directly attend to all other time steps, identifying relevant historical information for prediction. By processing the entire time series in parallel, the transformer architecture facilitates significant computational speedups compared to sequential models. The probabilistic forecasting capability is achieved by modeling the distribution of future values, rather than simply predicting point estimates, providing uncertainty estimates alongside the forecasts.
Chronos-2 utilizes two distinct attention mechanisms to improve forecasting accuracy. Time Attention focuses on capturing sequential dependencies within the historical demand of each individual good, allowing the model to learn patterns like seasonality and trends. Complementing this, Group Attention explicitly models the relationships between different goods purchased together, recognizing that demand for one item can influence the demand for others within a consumer’s typical bundle. This inter-dependency modeling is achieved by treating the set of goods as a collection and applying attention to learn the correlations between them, thereby improving the overall forecast by leveraging cross-product information.
LoRA, or Low-Rank Adaptation, is implemented to mitigate the computational expense associated with fine-tuning large transformer models like Chronos-2. This technique freezes the pre-trained model weights and injects trainable low-rank matrices into each layer of the Transformer architecture. By decomposing the weight update matrices into lower-dimensional representations, LoRA significantly reduces the number of trainable parameters – typically by over 90% – without substantially impacting performance. This reduction in trainable parameters lowers both the computational resources required for fine-tuning and the storage space needed for adapted models, enabling efficient experimentation and deployment on resource-constrained hardware.
Chronos-2 achieves effective modeling of complex demand patterns through the integration of Time and Group Attention mechanisms within a transformer architecture. These attention layers allow the model to capture both sequential temporal relationships and interdependencies between items, which is critical for forecasting in bundled consumer goods. Simultaneously, computational tractability is maintained via Low-Rank Adaptation (LoRA). LoRA significantly reduces the number of trainable parameters during fine-tuning, minimizing computational cost and memory requirements without substantial performance degradation. This combination facilitates efficient training and deployment of the model, enabling scalability to large datasets and complex forecasting scenarios.
Validating the Algorithm: From Synthetic Rationality to Real-World Choice
Synthetic data generation was employed to establish a ground truth for evaluating the model’s predictive capabilities regarding rational consumer choices. This process involved constructing datasets based on the principles of utility maximization, wherein consumers are assumed to make decisions that maximize their perceived benefit. The Generalized Axiom of Revealed Preference (GARP) was then applied to ensure the generated data adhered to consistent and rational consumer behavior. This allows for a controlled environment to assess the model’s ability to accurately forecast choices that align with established economic principles, independent of noise or confounding factors present in real-world datasets.
Fine-tuning the Chronos-2 model on synthetically generated data resulted in significant reductions in bundle prediction error across multiple forecast horizons. Specifically, errors decreased by 17-18% at horizons of 5, 10, and 15 periods, and by 31% at a horizon of 1 period, when compared to the performance of the model in a zero-shot configuration. These improvements demonstrate the value of the synthetic data in calibrating the model for predicting rational consumer choices and establishing a quantifiable performance baseline.
Model generalizability was assessed utilizing data from the Ahn et al. experiment, which presents a real-world portfolio-choice scenario involving consumer selections from a set of hypothetical product bundles. This experiment provided an independent dataset, distinct from the synthetic data used for initial training and validation, allowing for evaluation of the model’s performance on actual consumer behavior. The dataset comprises choices made by participants presented with varying bundle options, offering a benchmark for assessing the model’s ability to predict choices in a non-synthetic environment and to extrapolate learned patterns to new data distributions.
Analysis of portfolio-choice experiment data utilized the Critical Cost Efficiency Index (CCEI) to quantify consumer rationality, with higher CCEI values indicating more rational decision-making. Correlation analysis revealed a coefficient of 0.456 between the CCEI and bundle fitness – a metric representing the model’s predicted likelihood of selection. This positive correlation suggests a moderate, statistically significant relationship between the degree of consumer rationality, as measured by CCEI, and the accuracy of the model’s bundle predictions; however, it also indicates that other factors beyond pure rationality contribute to choice behavior.
Beyond Prediction: The Implications of Modeling Human Choice
The integration of established economic principles with the capabilities of modern machine learning presents a significant advancement in demand forecasting. This approach moves beyond purely data-driven predictions by incorporating understandings of rational consumer behavior, such as price sensitivity and substitution effects. The resulting models aren’t simply extrapolating past trends, but are instead building predictions on a foundation of economic theory, allowing them to better interpret market signals and generalize to novel situations. Consequently, forecasts become more robust, particularly during periods of disruption or when faced with limited historical data – ultimately providing businesses with a more reliable tool for inventory management, resource allocation, and strategic planning.
The predictive power of this forecasting model stems from its foundation in economic rationality – the assumption that consumers make consistent, logical choices to maximize their satisfaction. By incorporating this principle, the model isn’t simply memorizing past data, but rather learning the underlying drivers of demand. This allows it to extrapolate beyond the observed dataset and accurately predict responses to novel situations – new products, unexpected price changes, or shifts in market conditions. Unlike purely data-driven machine learning approaches prone to overfitting, this framework exhibits superior generalization capabilities, delivering robust and reliable forecasts even when confronted with previously unseen data, ultimately improving decision-making in dynamic environments.
The established forecasting framework possesses considerable adaptability, allowing for the incorporation of nuanced demand drivers beyond core economic principles. Researchers anticipate that integrating variables like seasonal trends – capturing predictable fluctuations throughout the year – and promotional activities, which demonstrably influence consumer purchasing, will significantly refine predictive accuracy. Furthermore, the model can be augmented to account for external events, ranging from localized disruptions like weather events to broader economic shifts or even viral social media trends. These additions, implemented through careful feature engineering and model retraining, promise a more robust and responsive demand forecasting system capable of navigating the complexities of real-world markets and delivering increasingly precise predictions.
Investigations are shifting towards a more nuanced understanding of consumer decision-making, acknowledging that perfect rationality is rarely observed in practice. Future models aim to incorporate the concept of bounded rationality, recognizing the cognitive limitations and informational constraints that influence choices. This entails exploring how consumers utilize simplifying heuristics, exhibit biases, and make ‘good enough’ rather than optimal decisions. By integrating these behavioral insights, researchers intend to develop forecasting tools that are more robust and accurate, particularly when faced with incomplete data or unpredictable market shifts. This progression seeks to move beyond idealized economic assumptions, creating a more realistic and powerful framework for anticipating consumer demand and responding effectively to evolving market dynamics.
The pursuit of accurate forecasting, as demonstrated by this work with Chronos-2 and revealed preference, reveals a fascinating truth about human behavior. Even when agents don’t adhere to strict rationality-a common occurrence in real-world economic activity-modeling their choices as if they were maximizing utility can yield surprisingly effective results. This aligns with a broader understanding that predictive power often stems not from capturing perfect logic, but from identifying consistent patterns, however flawed. As René Descartes observed, “Doubt is not a pleasant condition, but it is necessary to a clear understanding.” The researchers, by acknowledging the imperfections in consumer behavior, have crafted a method that improves forecasting by embracing the inherent ‘doubt’ in economic modeling and building from there, demonstrating that even flawed algorithms can approximate human decision-making.
What Lies Ahead?
The pursuit of rationality, even as a modeling assumption, remains a curious exercise. This work, by anchoring forecasting to the illusion of utility maximization, achieves improvements despite the acknowledged imperfections of human choice. The market, after all, is merely a barometer of collective mood, and mood rarely adheres to axiomatic consistency. Future iterations will likely focus on relaxing the GARP constraints – acknowledging that revealed preference is often a post-hoc justification, not a guiding principle.
A critical challenge lies in scaling this approach. The generation of synthetic data, while effective, introduces another layer of model dependence. How does one validate the simulator of the irrational agent? The field needs to consider incorporating behavioral biases directly into the data generation process, rather than attempting to smooth them away. Perhaps a more fruitful path involves explicitly modeling the sources of irrationality – loss aversion, hyperbolic discounting, and the myriad cognitive shortcuts that define human decision-making.
Ultimately, the value of this work isn’t in creating perfectly rational forecasts, but in building models that are robustly imperfect. Rationality is a rare burst of clarity in an ocean of bias. The goal shouldn’t be to eliminate the waves, but to navigate them with a little more understanding – and a little less faith in idealized assumptions.
Original article: https://arxiv.org/pdf/2603.23993.pdf
Contact the author: https://www.linkedin.com/in/avetisyan/
See also:
- Top 20 Dinosaur Movies, Ranked
- 20 Movies Where the Black Villain Was Secretly the Most Popular Character
- 25 “Woke” Films That Used Black Trauma to Humanize White Leads
- Silver Rate Forecast
- Spotting the Loops in Autonomous Systems
- Gold Rate Forecast
- Celebs Who Narrowly Escaped The 9/11 Attacks
- From Bids to Best Policies: Smarter Auto-Bidding with Generative AI
- 22 Films Where the White Protagonist Is Canonically the Sidekick to a Black Lead
- Can AI Lie with a Picture? Detecting Deception in Multimodal Models
2026-03-26 12:59