📊 statistics

On the retraining frequency of global models in retail demand forecasting

This paper demonstrates that less frequent retraining of global forecasting models across various machine learning and deep learning architectures maintains predictive accuracy while significantly reducing computational costs and energy consumption, offering a more sustainable and efficient approach to large-scale retail demand forecasting.

Original authors: Marco Zanotti

Published 2026-01-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Marco Zanotti

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you run a massive grocery store with thousands of products. Every day, you need to guess how many apples, toothbrushes, or cereal boxes people will buy so you don't run out or waste money on unsold stock. To do this, you use a "Global Model"—a super-smart computer brain that learns from the sales history of all your products at once, rather than trying to learn a separate rule for every single item.

For a long time, the standard advice was: "Keep this computer brain constantly updated." Every time a new sale happens, you should retrain the model immediately to keep it sharp. The paper by Marco Zanotti challenges this habit. He asks: Do we really need to retrain this model every single day, or can we let it rest a bit?

Here is the breakdown of his findings using simple analogies:

1. The "Constant Training" vs. "Periodic Check-up"

Think of the forecasting model like a student studying for a big exam.

The Old Way (Continuous Retraining): The student studies a little bit every single hour, every day. They are constantly re-reading the textbook the moment a new page is printed. This keeps them very up-to-date, but it's exhausting, expensive (in terms of time and energy), and they might get burned out.
The New Way (Periodic Retraining): The student studies hard for a week, then takes a break. They only go back to the textbook once a month to update their notes with the latest chapters.

The Paper's Discovery: The study found that the student who studies once a month performs just as well (or sometimes even better) on the exam than the one who studies every hour. The "Global Model" is smart enough to learn the general patterns of the store (like "people buy more ice cream in summer") without needing to be re-taught every single day.

2. The Cost of "Green" Efficiency

Retraining these models isn't free. It requires massive computer power, which uses electricity and creates carbon emissions (like driving a car).

The Analogy: Imagine the computer is a giant factory machine. Running it 24/7 to retrain the model every day costs a fortune in electricity.
The Result: By switching from "daily updates" to "monthly updates," the study found that companies could cut their computing costs by 75%. If they stopped retraining entirely and just used the model once, they could save 90% of the cost.
The Catch: For the most common type of prediction (guessing the exact number of items sold), the accuracy didn't drop at all. The model stayed sharp even while it "rested."

3. Machine Learning vs. Deep Learning

The paper tested two types of "brains":

Machine Learning Models: These are like efficient, practical workers. They are great at following rules and patterns. The study found that these workers benefit the most from taking breaks. As the data gets bigger, these models save the most money and energy when retrained less often.
Deep Learning Models: These are like genius artists who can see very complex, hidden patterns. They are powerful but require a lot of energy to run. While they also save money when retrained less often, they don't save quite as much as the practical workers, and they seem to hit a "ceiling" where taking longer breaks doesn't help them save much more time.

4. The "Safety Net" (Probabilistic Forecasting)

Sometimes, you don't just want to know how many apples will sell; you want to know the range of possibilities (e.g., "We will sell between 100 and 150 apples, but maybe up to 200"). This is called probabilistic forecasting.

The Finding: If you need these "safety net" predictions, the model does need to be updated a little more often than for simple guesses. However, even here, updating it once every two weeks is usually enough. You don't need to do it daily.

The Bottom Line

The paper argues that the old rule of "always keep the model fresh" is outdated and wasteful.

For Point Forecasts (Exact numbers): You can retrain your model once a month (or even less) and get the same accuracy as daily updates, while saving a massive amount of money and energy.
For Safety Stock (Ranges): You might need to update it every two weeks, but still far less than daily.

In short: Your global forecasting model is like a well-trained athlete. It doesn't need to run a sprint every single hour to stay in shape. It can run a marathon, rest for a few weeks, and still win the race. By letting it rest, businesses can save money and help the environment without losing their edge in predicting demand.

Technical Summary: On the Retraining Frequency of Global Models in Retail Demand Forecasting

Problem Statement

In the realm of retail demand forecasting, organizations face a critical trade-off between forecasting accuracy, computational efficiency, and sustainability. While Global Forecasting Models (GMs)—which fit a single model across an entire dataset of time series (cross-learning)—have gained prominence for their ability to leverage shared dynamics and reduce computational overhead compared to Local Models (LMs), the industry standard often involves continuous model retraining. This practice assumes that frequent updates are necessary to adapt to changing patterns. However, continuous retraining incurs substantial computational costs, energy consumption, and carbon emissions, particularly for large-scale applications involving millions of SKUs. The paper investigates whether this frequent retraining is truly necessary for maintaining the predictive performance of global models or if less frequent strategies can offer a more sustainable and cost-effective alternative without sacrificing accuracy.

Methodology

The study employs a rigorous experimental design to evaluate the impact of retraining frequency on global forecasting models.

Datasets: The analysis utilizes two comprehensive retail forecasting datasets:
- M5 Competition: Daily unit sales data for 30,490 SKU-store combinations from Walmart (2011–2016).
- VN1 Forecasting Challenge: Weekly sales data for 15,053 products from 328 warehouses across 46 retailers (2020–2024).
- The experiments focus on the most disaggregated level (SKU × Store) to maximize the potential benefits of reduced retraining.
Models: Ten distinct global forecasting methods were evaluated, comprising:
- Machine Learning (5 models): Linear (Pooled) Regression, Random Forest (RF), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LGBM), and Categorical Boosting (CatBoost).
- Deep Learning (5 models): Multi-Layer Perceptron (MLP), Long Short-Term Memory (LSTM), Temporal Convolutional Networks (TCN), Neural Basis Expansion Analysis for Time Series (NBEATS), NBEATS with Exogenous variables (NBEATSx), and Neural Hierarchical Interpolation for Time Series (NHITS).
Retraining Scenarios: The study defines a retrain window $r$ (the number of new observations before retraining). Scenarios range from:
- Continuous Retraining ( $r=1$ or $r=7$ for daily data): The benchmark for maximum accuracy and cost.
- Periodic Retraining: Intermediate intervals (e.g., weekly, monthly).
- No Retraining ( $r=T$ ): The model is fitted once on the initial training set and used for the entire test horizon.
Evaluation Framework:
- Validation: Rolling origin evaluation with an expanding window approach to simulate real-world business conditions.
- Metrics:
  - Point Forecasting: Root Mean Squared Scaled Error (RMSSE).
  - Probabilistic Forecasting: Scaled Multi-Quantile Loss (SMQL) across 14 quantiles.
  - Efficiency: Computing Time (CT) in seconds, directly correlated to monetary costs via cloud pricing models.
- Statistical Analysis: Friedman-Nemenyi tests were used to verify statistical significance across scenarios.

Key Results

1. Forecast Accuracy vs. Retraining Frequency

Point Forecasting: The study finds that less frequent retraining does not harm, and in some cases slightly improves, point forecast accuracy. For both M5 and VN1 datasets, the RMSSE profiles remained stable across retraining frequencies. Even in the "no retraining" scenario, performance degradation was negligible (less than 5%).
Probabilistic Forecasting: Results were slightly more nuanced. For the M5 dataset, accuracy (measured by SMQL) showed a slight decline as retraining frequency decreased, though the difference remained within 5-6 percentage points. For the VN1 dataset, an almost convex relationship was observed, where low retraining scenarios initially improved performance before deteriorating at higher intervals.
Model Differences: Machine learning models generally maintained or improved accuracy with less frequent retraining as data frequency increased, whereas deep learning models showed more stability but less marginal gain in this specific context.

2. Computational Cost and Efficiency

Exponential Reduction: Computing Time (CT) decreases exponentially as the retraining interval $r$ $r$ increases.
- Moving from continuous to monthly retraining reduced CT by approximately 75%.
- The "no retraining" scenario reduced CT by up to 90% compared to the baseline.
Cost Implications: Using a hypothetical large retailer scenario (200,000 SKUs, 5,000 stores), the study estimates that continuous retraining costs approximately $750,000 annually for daily data. Switching to a "no retraining" strategy could lower this to ~$250,000, representing savings of over 60%. For weekly data, savings reached 90%.
Model Selection: Machine learning models coupled with less frequent retraining strategies proved to be a marginally better choice for cost reduction as data frequency increased, outperforming deep learning models in terms of cost-efficiency gains.

3. Optimal Retraining Intervals

The analysis identified "optimal" retraining frequencies based on accuracy:

Point Forecasting: Optimal intervals were found to be between 3–4 weeks (approx. 21 observations) for M5 and 8–10 weeks for VN1.
Probabilistic Forecasting: Optimal intervals were shorter (approx. 2 weeks for M5, 4 weeks for VN1), suggesting that uncertainty estimation may require slightly more frequent updates than point estimates.

Key Contributions

Comprehensive Empirical Study: This is the first study to systematically analyze the relationship between retraining frequency and forecast accuracy across 10 diverse global models (ML and DL) using real-world retail datasets, covering both point and probabilistic forecasting.
Quantification of Trade-offs: The paper quantifies the trade-off between accuracy and sustainability, demonstrating that significant computational cost reductions (up to 90%) can be achieved with negligible impact on predictive performance.
Practical Guidelines: The study provides actionable recommendations for practitioners, suggesting that periodic retraining (e.g., monthly) offers a superior balance between accuracy and cost compared to the conventional continuous retraining paradigm.

Significance and Claims

The paper challenges the conventional wisdom that frequent or continuous retraining is essential for maintaining the high predictive performance of global forecasting models. The authors claim that global models, by learning shared dynamics across multiple time series, are robust enough to remain effective without constant updates, particularly in contexts with relatively stable demand patterns.

The significance of these findings lies in their dual impact:

Economic: Organizations can achieve substantial financial savings by optimizing retraining schedules without compromising forecast quality.
Environmental: By reducing the frequency of model training, organizations can significantly lower energy consumption and carbon emissions, aligning with the principles of "Green AI."

The study concludes that shifting from a continuous retraining paradigm to a periodic approach allows businesses to optimize their forecasting pipelines, achieving relevant cost savings while ensuring high predictive performance. The authors note that while their findings are robust for stable retail demand, future research should explore adaptive retraining strategies triggered by concept drift detection and extend these analyses to other domains and model architectures.