Deep Learning vs. Statistical Models for Multi-Horizon Price Forecasting of Second-Hand Electronics: A Systematic Benchmark

arXiv:2610.10727 · q-fin.CP, cs.LG · Submitted 2026-10-07 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Deep Learning vs. Statistical Models for Multi-Horizon Price Forecasting of Second-Hand Electronics".

Jane: Forecasting resale prices of used electronics is critical for subscription-based platforms where pricing errors translate directly into risk.

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we’ve looked at the title of "Deep Learning vs. Statistical Models for Multi-Horizon Price Forecasting of Second-Hand Electronics: A Systematic Benchmark." It immediately signals the core conflict they are testing right there.

Jane: That’s right; it sets up a direct comparison between established statistical methods and newer deep learning architectures on the specific task of predicting used electronics prices over multiple time steps.

Lu: The authors are Mateusz Buczynski, Michał Woźniak, Konrad Kaczynski, Anna Wroblewska, and Sebastian Kukc. They are researchers from the Faculty of Economic Sciences at the University of Warsaw and also from the Faculty of Mathematics and Information Science at Warsaw University of Technology.

Meng: So it’s a collaboration between economics faculty and pure computer science faculty; that mix suggests they're looking for a very solid, mathematically sound approach to solving this practical forecasting problem.

Tom: And the implication is pretty direct: they are providing the first systematic multi-horizon benchmark specifically for used electronics price forecasting. This isn't just another study; it’s establishing a new standard for how we test models in this domain #pg1

Jane: It’s about moving beyond just treating pricing as a simple static regression problem and giving it that necessary temporal dimension, which is what the authors say is missing in prior work on second-hand goods.

Lu: The motivation here is clearly rooted in addressing those four specific gaps they identified earlier, especially the lack of any time-series benchmark for used electronics price prediction #pg3

Meng: So when you put it all together, this paper isn't just predicting prices; it's building the framework that others will use to decide which forecasting tool is best suited for a domain that is complex and volatile like used goods.

The paper's summary: Tom: Now let’s get into what the paper actually summarizes in detail. Essentially, they are using a large-scale dataset of daily price listings from Polish online marketplaces spanning from January two thousand twenty-two to March two thousand twenty-five <ref:2610.10727#pg1>.

Jane: They collected over one hundred smartphone and laptop models, and they used this data to test eleven different models across six horizons ranging from one day up to three hundred sixty-five days #pg2

Lu: The summary highlights that the method for creating the time series involves taking those daily listings, matching abstract item names to specific listings using keyword-based scoring, fuzzy string matching, Named Entity Recognition models, and text embeddings.

Meng: That’s a pretty detailed data pipeline just to get that clean price time series ready for modeling; it shows how much effort goes into getting usable data from messy sources.

Tom: And the summary stresses the evaluation strategy: they use three protocols—trajectory fitness, one-shot terminal accuracy, and cross-horizon model reuse—to check different aspects of a model’s performance.

Jane: It's important because it shows that they aren't relying on just one metric to judge success; they are checking trajectory accuracy, endpoint accuracy at the end point, and how well a model can be transferred across different time windows #pg2

Lu: They also detail the hyperparameter analysis they did, showing how they looked at the standard deviation of core hyperparameters like learning rate, batch size, maximum training steps, and input window size #pg1

Meng: That’s good because it gives us insight into what makes a model stable or unstable during the training process; it shows which configurations are more reliable to deploy in production.

The paper's improvements: Tom: So when we look at the specific improvements the authors suggest, they focus heavily on making their framework more robust for real-world deployment. They highlight that N-BEATS and N-HiTS are simply more robust to hyperparameter selection than transformer architectures.

Jane: That’s because they found that while transformer models can be fragile, especially when predicting long horizons, these two models handle configuration changes much better #pg1

Lu: And they pointed out the practical improvement of using a single model; the finding here is that one N-BEATS model trained for three hundred sixty-five days maintains competitive accuracy at all those shorter horizons without having to be retrained #pg1

Meng: That’s a big win for operational simplicity because it means you don't have to manage and maintain a whole suite of horizon-specific models just to cover the short term.

Tom: And on the error front, they quantified the improvement as deep learning reducing forecasting error by forty-three percent versus the best statistical baseline when looking at long term predictions #pg1

Jane: That means that for subscription platforms, using these deep learning methods could translate directly into a substantial reduction in the risk of pricing errors.

Lu: They also mentioned that residuals at long horizons being negative on average is an observation they made about N-BEATS and N-HiTS; it suggests these models are consistently underestimating the depreciation trend #pg1

Conclusion: Tom: So to wrap up, this paper, "Deep Learning vs. Statistical Models for Multi-Horizon Price Forecasting of Second-Hand Electronics: A Systematic Benchmark," shows that deep learning has a definite edge when looking beyond thirty days in this domain.

Jane: The overall implication is that while statistical models can approximate the short term dynamics reasonably well, they really struggle to capture the long-term nonlinearities and structural shifts present in the used electronics market #pg1

Lu: The main empirical finding is that N-BEATS consistently achieves the lowest forecasting error across all horizons beyond thirty days at eight point five one percent for three hundred sixty-five days compared to fourteen point nine four percent for the best statistical baseline #pg2

Meng: The practical implication is that we have a solid tool—N-BEATS, and they found it's robust and can even cover multiple horizons from one training run #pg1

Lalam: This research helps us understand how to build more resilient AI systems by clearly showing what works when dealing with high volatility in pricing data like this.

Tom: So that’s our discussion on this paper for today, and we're ready to move on to whatever interesting stuff is waiting in the queue for next time.

Mateusz Buczynski, Michał Woźniak, Konrad Kaczyński, Anna Wroblewska, Sebastian Kuk

Faculty of Economic Sciences, University of Warsaw · Faculty of Mathematics and Information Science, Warsaw University of Technology

q-fin.CP, cs.LG

Submitted: 2026-10-07

Updated: 2026-10-07

Comments: 41 pages, 7 figures, 13 tables

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 83/100

The gist: Forecasting resale prices of used electronics is critical for subscription-based platforms where pricing errors translate directly into risk.

Key concepts

Multi-Horizon Forecasting: This involves predicting a price not just for tomorrow, but across several future time steps simultaneously. The study tested models to see which could best predict prices over short (1 day) and long (365 days) periods, which is crucial for planning.
N-BEATS Model: N-BEATS is a specific deep learning architecture used for forecasting. It was found to be the most accurate model in this research, achieving the lowest error rate when predicting used electronics prices over long timeframes like a year.
Evaluation Protocols: The researchers used three distinct ways to test models: Protocol 1 checks if a model fits the expected price path (trajectory fitness), Protocol 2 checks accuracy at a single future point, and Protocol 3 tests if one model trained for one horizon can be reused for others.
MAPE: Mean Absolute Percentage Error is the main metric used to measure forecast accuracy. It tells researchers, on average, by what percentage the model's predictions deviate from the actual observed prices in the used electronics market.

Terminology

Benchmark — 1× in this paper · explained in 2 episodes

Summary

Forecasting resale prices of used electronics is critical for subscription-based platforms where pricing errors translate directly into risk. The gist: N-BEATS consistently achieves the lowest forecasting error across all horizons beyond 30 days, reaching MAPE = 8.51% at the 365-day horizon compared to 14.94% for the best statistical baseline

Research Gap and Motivation

The paper addresses a significant research gap by presenting the first multi-horizon benchmark of statistical and deep learning forecasting models for used electronics price prediction Prior research has largely focused on structured markets or treated pricing as a static tabular regression task without a temporal dimension The study identifies four specific, unaddressed gaps in the literature, including the lack of a time-series benchmark for used electronics.

Systematic Benchmark and Models

The study establishes a systematic multi-horizon benchmark by evaluating eleven models across six horizons from 1 to 365 days These models span classical statistical methods (ARIMA, ETS, Theta), recurrent and convolutional networks (LSTM, TCN), and modern deep architectures (N-BEATS, N-HiTS, TFT, PatchTST, Informer). The evaluation utilizes three complementary protocols to assess trajectory fitness (Protocol 1), one-shot endpoint accuracy (Protocol 2), and cross-horizon transfer (Protocol 3).

Key Findings on Model Performance

The empirical findings reveal distinct performance patterns across different horizons. At short horizons (1–7 days), all models converge near 0.72% MAPE, with the naive baseline remaining competitive As the forecast horizon increases to 365 days, deep learning models begin to show a distinct advantage, with N-BEATS achieving the lowest MAPE beyond 30 days at 8.51% compared to 14.94% for the best statistical baseline. Conversely, recurrent models like LSTM and TCN degrade sharply beyond 90 days, with LSTM reaching a MAPE of approximately 0.372 at the 365-day horizon.

Evaluation Protocols and Robustness

The evaluation framework is meticulously specified to meet business requirements, utilizing metrics such as Mean Absolute Percentage Error (MAPE) as the primary criterion. The study also incorporates a hyperparameter robustness analysis, which quantifies sensitivity per model and horizon, providing actionable guidance on deployment risk. The results show that N-BEATS and N-HiTS are more robust to hyperparameter selection and better suited for consistent deployment compared to transformer-based architectures.

Operational Insights

A practically important result from Protocol 3 is that a single N-BEATS model trained for the 365-day horizon maintains competitive accuracy at all shorter horizons, without retraining. This finding suggests that practitioners can deploy a single production model rather than a family of horizon-specific models, reducing operational complexity significantly. Error source analysis indicates that for N-BEATS and N-HiTS, residuals at long horizons are increasingly negative on average, consistent with models failing to fully capture the structural downward depreciation trend.

The study concludes that while statistical models can approximate short-term dynamics well, they struggle to capture long-term nonlinearities and structural shifts present in the used electronics market. The central empirical finding is that N-BEATS consistently achieves the lowest forecasting error across all horizons beyond 30 days, reaching MAPE = 8.51% at the 365-day horizon compared to 14.94% for the best statistical baseline. The gist>

The paper introduces a real-world used electronics price dataset - the first of its type constructed specifically for multi-horizon forecasting. It provides the first systematic benchmark of eleven models across six horizons in this domain. It proposes a three-protocol evaluation framework that separates trajectory accuracy, endpoint accuracy, and cross-horizon generalization. It delivers the first hyperparameter robustness analysis for automated pricing systems in this market. The gist>

How it works

The dataset consists of scraped daily listings from social media and publicly available offer sites in Poland that facilitate the sale of used goods To ensure accurate price tracking, a system was developed to match abstract item names to specific listings using keyword-based scoring, fuzzy string matching, Named Entity Recognition (NER) models, and text embeddings. Daily price listings for each abstract item are aggregated using a simple mean to form a structured time series

Model Selection and Training

Model selection follows a generational structure, including classical statistical methods (ARIMA, ETS, Theta), recurrent and convolutional neural networks (LSTM, TCN), and modern deep architectures (N-BEATS, N-HiTS, TFT). Each model receives a univariate historical time series representing the daily smoothed price of a specific electronic item Deep learning models are configured as direct multi-output models that predict all h future time steps simultaneously in a single forward pass.

Evaluation Strategy

The evaluation employs three complementary protocols to distinguish between trajectory fitness (Protocol 1), one-shot terminal accuracy (Protocol 2), and cross-horizon model reuse (Protocol 3). Temporal Double Cross Validation based on the Forward Chaining approach is used for evaluation, allowing models M within a specified forecast horizon h across w out of sample (OOS) time windows.

Hyperparameter Analysis

Variance analysis across top configurations was conducted to compute the standard deviation of core hyperparameters—learning rate, batch size, maximum training steps, and input window size This analysis showed that N-BEATS and N-HiTS are more stable under different configurations than transformer-based architectures.

Limitations

Several limitations of this study should be acknowledged, including a single-market scope drawn exclusively from Polish online marketplaces The dataset is biased toward high-end and Apple-branded devices. Smoothing attenuation substantially reduces apparent daily volatility, meaning reported MAPE values reflect the error on smoothed series, not on raw listing prices. No external covariates are included, such as macroeconomic indicators or search trends The study produces only point forecasts and lacks probabilistic forecasting, which would be more informative for subscription pricing decisions.

The paper's findings confirm that second-hand electronics price series are not inherently harder to forecast than structured financial series, provided that appropriate deep learning architectures are employed. The gist>

Model Selection and Training

Improvements for AI systems

Bold header: Multi-horizon generalization capability. This system can achieve A single one-year model generalizes to all shorter horizons without retraining, meaning a model trained for 365 days can be deployed across all required forecasting horizons (1 to 365 days).
Bold header: Superior long-term error reduction. The N-BEATS model achieves best accuracy from 30 days to one year ahead, reducing forecasting error by 43% versus best statistical baseline.
Bold header: Robustness against hyperparameter instability. The analysis reveals that NBEATS and N-HiTS are more robust to hyperparameter selection and better suited for consistent deployment compared to transformer-based architectures, which are noted as being fragile, especially at longer horizons.

Abstract

Forecasting resale prices of used electronics is critical for subscription-based platforms where pricing errors translate directly into risk. Unlike structured financial markets, second-hand electronics exhibit high volatility, sparse listing histories, and non-normal price dynamics - yet no systematic time-series benchmark exists for this domain. This paper presents the first multi-horizon benchmark of statistical and deep learning forecasting models for used electronics price prediction. We use a large-scale dataset of daily price listings from Polish online marketplaces (January 2022 to March 2025, 100+ smartphone and laptop models) and evaluate eleven models across six horizons from 1 to 365 days, covering classical methods (ARIMA, ETS, Theta), recurrent and convolutional networks (LSTM, TCN), and modern deep architectures (N-BEATS, N-HiTS, TFT, PatchTST, Informer). Three complementary evaluation protocols assess trajectory fitness, one-shot endpoint accuracy, and cross-horizon transfer. N-BEATS achieves the lowest MAPE beyond 30 days, reaching 8.51% at 365 days versus 14.94% for the best statistical baseline - a 43% reduction. At short horizons (1-7 days), all models converge near 0.72% MAPE and the naive baseline remains competitive. A single N-BEATS model trained at 365 days generalizes to all shorter horizons, eliminating the need for horizon-specific models. N-BEATS and N-HiTS also demonstrate superior hyperparameter stability.