The State of Time Series Forecasting in 2025: From Classic Models to AI Titans
For decades, the path forward in time series forecasting seemed clear: a steady march from simple statistical models to ever more complex deep learning architectures. That path has been completely rerouted.
As of 2025, the field is in a state of profound transformation, an "architectural renaissance" where the old rules no longer apply.
The simple narrative of "more complexity equals more accuracy" has been challenged by a surprising resurgence of simpler models. At the same time, the arrival of massive, pre-trained foundation models, akin to the AI that powers ChatGPT, is set to redefine the forecasting landscape entirely.
This guide provides a comprehensive overview of the current state of time series forecasting. It’s a strategic playbook for practitioners and researchers, designed to help you navigate this dynamic and exciting new era. We'll explore the core models you need to know, the cutting-edge foundation models, and a practical framework for choosing the right tool for your specific challenge.
Part 1: The Great Reality Check, Why Simplicity Is Back in Style
The single most important lesson from recent forecasting research is a powerful validation of the "No Free Lunch" theorem: there is no universally superior model.
For a period, complex Transformer-based architectures were seen as the pinnacle of forecasting. But this assumption was challenged by a series of critical studies in 2024 and 2025. This research delivered a shocking result: simple, lightweight linear models could often outperform these sophisticated deep learning architectures, particularly on long-term forecasting benchmarks.
This "linear model challenge" forced a deep and necessary introspection within the research community. It revealed that many standard benchmarks penalized overly complex models that were prone to overfitting on data where the underlying statistical properties shifted between the training and testing periods.
The key takeaway is that classical statistical models remain formidable benchmarks. They are not just historical artifacts; they are often the most robust, interpretable, and cost-effective solution for a wide range of real-world problems.
This leads to the foundational principle of modern forecasting: Always start with strong baselines. Before deploying a complex neural network, you must prove it can consistently and significantly outperform these simple but powerful alternatives:
- A Naive Forecast: Predicting the last observed value (or the value from the previous season). For highly volatile series like stock prices, this is often impossible to beat.
- An ETS Model: A fast, robust, and automatic powerhouse for data with clear structural components.
- A SARIMA Model: The go-to for series with clear seasonality and the influence of known external drivers.
If your complex model can't beat these, its added cost, training time, and opacity are not justified.
Part 2: The Core Toolbox, A Deep Dive into Your Model Options
In today's diverse landscape, model selection requires a nuanced understanding of each architecture's strengths and weaknesses. Here’s a detailed breakdown of the core models that form the modern practitioner's toolbox.
The Statistical Bedrock: SARIMAX and ETS
These models are the foundation of forecasting, offering interpretability and robust performance.
- SARIMAX (Seasonal Autoregressive Integrated Moving Average with eXogenous Regressors): This powerful model captures a wide range of time series dynamics. Its strength lies in its interpretability; each parameter corresponds to a clear statistical property like trend, seasonality, or dependence on past values. The "X" for exogenous variables allows you to incorporate the impact of known external drivers, such as marketing campaigns or policy changes, making it invaluable for business and economic forecasting.
- Best for: Data with clear, stable trends and seasonality, especially when you need to understand the impact of external factors.
- ETS (Error, Trend, Seasonality): The ETS framework is a family of state-space models that provide a statistical foundation for exponential smoothing methods. ETS models are famously robust, computationally fast, and often serve as a difficult-to-beat baseline. They automatically handle various types of trend and seasonality, making them incredibly easy to apply.
- Best for: A strong, fast, and robust baseline for a wide variety of business time series.
The Decomposable Innovators: Prophet and NeuralProphet
Developed to make forecasting more accessible, these models focus on intuitive, decomposable components.
- Prophet: Developed by Facebook, Prophet is a decomposable model that represents a time series as a sum of a trend, seasonality, and holidays. Its greatest asset is its "human-in-the-loop" design. You can easily plot and inspect each component, making it exceptionally interpretable for business stakeholders. It's robust to missing data and shifts in trend, making it an excellent plug-and-play baseline. Its primary weakness is a lack of an autoregressive component, making it less suitable for short-term forecasts where local context is key.
- Best for: Quick, interpretable business forecasts where trend, seasonality, and holidays are the main drivers.
- NeuralProphet: Built on PyTorch, NeuralProphet was created to address the core limitations of the original. It keeps the same intuitive decomposable structure but adds a key innovation: an AR-Net, a simple neural network that models autoregressive effects. This provides the "local context" Prophet lacks, often improving short-term forecast accuracy. It can also model non-linear relationships with external variables. The trade-off is increased complexity and more hyperparameters to tune.
- Best for: Problems where Prophet's components are useful but short-term autoregressive patterns are also critical.
The Machine Learning Powerhouse: XGBoost
- XGBoost (eXtreme Gradient Boosting): A dominant algorithm for general tabular data, XGBoost can be powerfully applied to forecasting. To do so, the time series is transformed into a supervised learning problem using a "sliding window" approach, where lagged values and time-based features (like day of the week) become predictors. XGBoost excels at capturing complex, non-linear relationships and interactions between features. Its critical limitation, however, is that as a tree-based model, it cannot extrapolate trends. It can only predict values within the range seen during training, making it unsuitable for series with a continuing growth trend.
- Best for: Forecasting stationary or de-trended series with many features and complex, non-linear relationships.
Synthesis and Performance Matrix
To help you navigate these options, this table synthesizes the key characteristics of each model.
Synthesis and Performance Matrix
| Model | Type | Key Strengths | Key Weaknesses | Interpretability | Ideal Use Case |
|---|---|---|---|---|---|
| SARIMAX | Statistical | High interpretability; strong theoretical foundation; explicitly handles seasonality and exogenous variables. | Assumes linearity; struggles with complex non-linear patterns; requires careful parameter tuning. | High | Forecasting data with clear, stable trend/seasonality and known external drivers (e.g., economic data). |
| ETS | Statistical | Robust and often highly accurate baseline; computationally very fast; automatically handles various trend/seasonality types. | Primarily for univariate series; assumes linearity; less direct handling of exogenous variables. | High | A strong, fast, and robust baseline for a wide variety of business time series. |
| Prophet | Decomposable | Extremely intuitive and interpretable; robust to missing data and trend shifts; excellent "plug-and-play" baseline. | Lacks autoregressive component (local context); often less accurate than statistical models for short-term forecasts. | Very High | Quick, interpretable business forecasts where trend, seasonality, and holidays are the main drivers. |
| Neural Prophet | Hybrid (NN) | Adds autoregression (AR-Net) to Prophet, improving short-term accuracy; flexible PyTorch backend. | More hyperparameters to tune; can be less stable than Prophet; may still be outperformed by simpler models like ETS. | High | Problems where Prophet's components are useful but short-term autoregressive patterns are also critical. |
| XGBoost | Machine Learning | Excellent at capturing complex non-linearities and feature interactions; highly scalable and efficient. | Cannot extrapolate trends; requires transforming data into a supervised problem. | Low | Forecasting stationary or de-trended series with many features and complex relationships. |
| Ensemble/Hybrid | Meta-Model | Typically yields the highest accuracy and robustness by combining strengths of diverse models. | Increased complexity in implementation and maintenance; can be computationally expensive. | Varies | Mission-critical applications where maximizing accuracy is the primary objective. |
Part 3: The New Frontier, Foundation Models Are Changing Everything
The most significant development shaping the future of TSF is the arrival of the foundation model. This paradigm shifts away from training a specialized model for each dataset. Instead, a single, massive model is pre-trained on a vast corpus of data, enabling it to perform a wide range of tasks with minimal or no task-specific training, a capability known as zero-shot learning.
Two distinct approaches are leading this revolution:
- Repurposing Large Language Models (LLMs): This approach aims to leverage the power of existing LLMs like GPT. Models like Time-LLM use clever techniques to "reprogram" a frozen LLM, teaching it to align the numerical data of a time series with its text-based understanding of the world.[1]
- Purpose-Built Time Series Foundation Models: This is currently the more dominant and successful approach. These models are trained from the ground up on massive corpora of time series data. They are designed to learn the fundamental "language of time" directly.
Leading purpose-built models include:
- Chronos (Amazon): A model that tokenizes time series values and trains a language model architecture on them, delivering remarkable zero-shot performance.
- TimesFM (Google): Trained on a colossal 100-billion-point dataset, TimesFM has been instrumental in proving that "scaling laws" (more data and bigger models lead to better performance) apply to time series.[2]
- Lag-LLaMA: A decoder-only Transformer that uses a unique lag-based tokenization strategy, allowing it to generalize to new time series with different frequencies.
- Moirai (Salesforce): A universal model designed to handle various data frequencies and covariates, and to generate probabilistic forecasts.[2]
The practical implications of foundation models are profound. They promise to dramatically simplify forecasting pipelines, democratize access to state-of-the-art models, and shift the role of the data scientist from model architect to data curator and prompt engineer.
Foundation Models Matrix
| Model (Source) | Base Architecture | Key Innovation / Approach | Primary Use Case |
|---|---|---|---|
| Time-LLM | Repurposed LLM | Reprograms a frozen LLM using lightweight, trainable input/output layers. | Leveraging existing LLMs for zero-shot forecasting. |
| Chronos (Amazon) | T5 Language Model | Tokenizes time series via simple scaling and quantization. | A simple, effective, and probabilistic foundation model for general-purpose zero-shot forecasting. |
| TimesFM (Google) | Decoder-Only Transformer | Trained from scratch on a massive (100B point) corpus of time series data. | High-performance zero-shot forecasting. |
| Lag-LLaMA | Decoder-Only Transformer | Uses a fixed set of frequency-based lags as input covariates, enabling generalization to unseen frequencies. | General-purpose univariate probabilistic forecasting. |
| Moirai (Salesforce) | Encoder Transformer | Universal architecture handling any data frequency and covariates with probabilistic outputs. | Universal forecasting, especially for complex scenarios with external variables. |
Part 4: Local vs. Cloud, Choosing Your Hardware
The choice of computing hardware is a critical, practical consideration that depends entirely on the model you select.
- Running Locally (Your Laptop): Classical models like SARIMAX, ETS, and Prophet are very lightweight and run perfectly on any modern laptop. They are ideal for initial exploration and smaller scale tasks. Machine learning models like XGBoost and NeuralProphet can also run locally, but performance will slow down with larger datasets.
- Running on the Cloud (AWS, GCP, Azure): For more demanding tasks, the cloud is essential.
- GPU Acceleration: Training NeuralProphet or custom deep learning models (LSTMs, Transformers) on large datasets becomes practical only with cloud GPUs.
- Inference with Foundation Models: While some smaller foundation models (like the tiny version of Chronos) can run on a powerful local machine, using their full-sized versions or deploying them for real-time predictions in a production environment requires a cloud endpoint. The models are large, and consistent, reliable performance is best managed in a cloud environment.
- Training Foundation Models: This is the exclusive domain of major tech corporations and large research institutions. Pre-training a model like TimesFM requires an immense investment in cloud infrastructure, using hundreds of specialized AI accelerators (GPUs/TPUs) for weeks at a time. For the vast majority of users, the only feasible path is using these pre-trained models, not building them from scratch.
Part 5: A Strategic Framework for Action, Your Playbook for 2025
Navigating this complex landscape requires a pragmatic and disciplined approach. The goal is not to find the single "best" model but the most appropriate model for your specific context.
Here is a step-by-step decision framework for your next forecasting project:
- Assess Data Predictability First: Before anything else, run a naive (no-change) forecast. If more complex models fail to show a significant and stable improvement, the series may be inherently unpredictable. In this case, focus on quantifying uncertainty rather than chasing marginal gains in accuracy.
- Define Interpretability Needs: Is it critical for stakeholders to understand why the forecast is what it is? If yes, prioritize interpretable models. Start with Prophet for its intuitive visualizations or ETS/SARIMAX for their clear statistical parameters.
- Characterize Data Patterns: Does the data have strong, stable seasonality and trend? Are there known external drivers (e.g., promotions, holidays)? SARIMAX is explicitly designed for this.
- Identify Non-Linearity: Does the series exhibit complex patterns that aren't simple trends or seasons? XGBoost is a powerful choice for this, but only if you have de-trended the data first.
- Consider Data Volume and Frequency: Is the data high-frequency (e.g., hourly) with large volumes? This is where deep learning models shine. NeuralProphet is a great starting point, balancing power and interpretability.
- Evaluate the Foundation Model Use Case: Are you facing a "cold start" problem with a new task and limited data? Do you need a powerful general baseline without weeks of training and tuning? This is the perfect time to try a zero-shot forecast from a model like Chronos or TimesFM.
- For Mission-Critical Accuracy, Go Hybrid: For your most important forecasts, the most robust solution will rarely be a single model. The state-of-the-art in applied forecasting is the ensemble or hybrid model. By strategically combining the predictions of a statistical model, a tree-based model, and a neural network, you can leverage the strengths of each while mitigating their individual weaknesses.
The Future is Now
The world of time series forecasting is more dynamic and complex than ever before. The linear progression from simple to complex has been replaced by a vibrant, cyclical ecosystem of diverse architectures.
The key to success in 2025 is not to master a single "state-of-the-art" model, but to cultivate a deep understanding of the entire toolbox. Be pragmatic. Be rigorous in your evaluation. And always, always start with the simple baselines. The future of forecasting belongs to those who can choose the right tool for the job.