The data science market is currently experiencing a surge in demand for machine learning across all sectors, sometimes even excessively. Sales forecasting is no exception, with machine‑learning–driven algorithms increasingly promoted as the new gold standard.
Certainly, the rise of highly performant machine learning models in recent years is undeniable. However, it remains essential to keep a critical distance from marketing narratives, no matter how attractive they may seem.
Understanding the scope and limitations of the approach, knowing when to use it based on actual business needs and organizational capabilities, and being able to interpret the results are all essential conditions to truly create value.
History of Machine Learning in Forecasting
Machine learning and artificial intelligence are often presented as revolutionary technologies, generating disproportionate excitement. Yet machine learning emerged as early as the 1950s and 1960s and gained momentum from the 1980s onward. Its use in forecasting dates back to the 1960s with decision trees and advanced significantly in the 1990s.
During this period, traditional statistical forecasting models—such as simple, double, and triple exponential smoothing, as well as autoregressive techniques—proved highly robust. These models are well‑suited to regular, long‑term sales histories. For intermittent or erratic demand, Croston’s method, developed in the 1970s, became an industry reference. Until just a few years ago, these classical approaches remained the most reliable options and often still outperform alternatives on stable time series.
It was only around 2016, when gradient boosting techniques matured for regression problems, that machine learning began consistently matching traditional forecasting performance. This shift was enabled by new technologies, broader access to cloud infrastructures handling massive datasets, and significant increases in computational power.
This opened the door to many innovations and explains the rapid growth of machine learning in forecasting since the early 2020s.
Recent Developments in Machine Learning
Since 1982, the Makridakis Competitions have gathered researchers worldwide to benchmark forecasting methods. For nearly 40 years, no classical model had been surpassed. Notably, in the 2020 edition, the seventeen machine learning models tested performed worse than the simplest Naive model – a surprising outcome.
Yet within a year, the landscape changed dramatically. New machine learning models integrating additional explanatory variables (“features”) and advanced mechanisms to avoid overfitting began outperforming all previous approaches. In retail, such features might include prices, historical trends, promotions, stockouts, or distribution channels. The quality and relevance of this enriched data significantly improve forecasting accuracy.
Since 2021, machine‑learning research has accelerated: larger datasets, heavier computational capabilities, international competitions, and massive community involvement. Feature‑rich models have rapidly improved in coherence and computation time, often overtaking classical methods.
Trade‑offs remain, however. These models are resource‑intensive, more vulnerable to overfitting, and heavily dependent on clean, well‑structured data. Poor‑quality data can degrade performance. They are also less robust, more unstable, and sometimes harder to interpret. Unlike classical models, which can often be applied immediately, advanced machine‑learning methods require expert configuration and ongoing supervision.
Deep Learning and the Emergence of Foundation Models
In parallel, deep learning has undergone rapid transformation through architectures like Transformers and self‑supervised learning strategies. These breakthroughs have enabled the rise of large‑scale foundation models capable of generalizing across many forecasting tasks by learning from massive and diverse datasets. As they scale, these models show emergent abilities, improving in reasoning, robustness, and multimodal understanding.
A striking example is Chronos‑2, a 120‑million‑parameter time‑series foundation model designed for zero‑shot forecasting. Supporting univariate, multivariate, and covariate‑informed tasks, it uses a T5‑inspired encoder and group attention mechanisms to capture dependencies across related series. Chronos‑2 achieves state‑of‑the‑art results on benchmarks such as fev‑bench and GIFT‑Eval while remaining computationally efficient.
Compared to classical machine‑learning approaches such as XGBoost, deep learning – and foundation models in particular – offers several advantages:
- automatic learning of complex, nonlinear temporal patterns
- native handling of multivariate interactions
- generalization to new tasks without retraining
Chronos‑2 illustrates this by modeling long‑range dependencies and external covariates directly, while XGBoost requires manual feature engineering, retraining, and struggles with high‑dimensional temporal structures.
While XGBoost remains a strong option for small, tabular, highly interpretable datasets, foundation models now dominate large‑scale, complex forecasting scenarios requiring flexibility, adaptability, and cross‑domain intelligence.
Adopting Machine Learning Wisely
The key is not to “use machine learning for the sake of using machine learning.” Skipping the foundational steps is counterproductive.
A pragmatic approach begins with:
- establishing clear business objectives
- understanding data requirements
- ensuring data quality and segmentation
- identifying whether machine learning is truly suited to the forecasting context
Once more advanced models are introduced, it is essential for organizations to ensure that results – often derived from complex algorithms – are interpretable and actionable. This fosters trust, adoption, and relevant monitoring aligned with business realities.
Machine learning does not benefit all contexts equally. Its value emerges only when the chosen model aligns with the specific needs of the targeted forecasting scenario. Between simple statistical models and sophisticated neural‑network architectures lies a wide range of options. Success depends on rigorous data preparation, methodological discipline, and progressive implementation.
In this landscape, solutions such as Colibri can provide structured guidance, helping organizations identify the right level of sophistication and adopt machine learning at a pace that aligns with their data maturity and operational goals.
Also read : Multi-Module S&OP Software | 5 Reasons to Adopt
From algorithms to real-world impact
While the debate around machine learning in forecasting often focuses on models and performance, real value also depends on how these technologies are deployed and adopted in practice. Beyond algorithms, organizations need structured approaches and expert support to translate data into actionable decisions. This is why Colibri has progressively built a network of specialized partners to complement its platform. An early example is the partnership with ALOER, introduced in 2018, which highlights how combining advanced forecasting capabilities with supply chain expertise helps companies move from theoretical accuracy to operational impact.