How to choose a forecasting platform or library

Every tool in this category is an open-source Python library — there's no vendor platform to license here, and no per-seat pricing to negotiate. The buying decision is really an engineering one: which library's model of the problem, level of automation, and computational approach matches how many series you're forecasting and how much your team wants to control the modeling itself. That makes this category unusual — read it as a guide to picking the right tool for your engineers, not a guide to picking a vendor.

How many series are you forecasting?

Scale is the first fork, because it changes which libraries are even practical:

• One series, or a handful, with clear seasonality. Prophet was built by Meta's data science team specifically to make a reasonable forecast easy with minimal tuning — it fits trend, yearly/weekly/daily seasonality and holiday effects, and is deliberately forgiving of missing data and outliers. It's aimed at analysts who need a quick, interpretable forecast, not machine-learning specialists.
• Thousands to millions of series at once. Nixtla's open-source libraries (StatsForecast, MLForecast, NeuralForecast) are engineered to fit huge numbers of individual series in parallel rather than one at a time — a different engineering problem than forecasting a single series well.
• A large panel of related series with shared structure — many stores, many SKUs, many sensors. PyTorch Forecasting's TimeSeriesDataSet abstraction is built specifically to encode multiple related series, static and time-varying covariates, and variable-length histories into batches a neural network can train on.

How much modeling control do you want?

• Maximum automation. AutoGluon's AutoGluon-TimeSeries module trains and ensembles a pool of statistical, gradient-boosting and deep-learning forecasters — including its own pretrained Chronos foundation model — and automatically selects or blends the best performers, so a strong baseline takes a few lines of code and no model-selection expertise. AWS now points users here, paired with SageMaker, as its recommended forecasting path.
• A no-training hosted model. Nixtla's TimeGPT is a pretrained, hosted foundation model accessed via API for teams that want a forecast without training anything, supporting fine-tuning, external variables and prediction intervals — this is Nixtla's one paid, quote-based product, distinct from its free open-source libraries.
• Benchmark many model types against each other. Darts's defining idea is a single, consistent fit/predict API across dozens of models — classical statistical, gradient-boosting, deep-learning architectures and pretrained foundation models — so analysts can swap models, backtest several side by side, or ensemble without rewriting data-handling code.
• scikit-learn-style rigor. sktime brings scikit-learn's familiar estimator interface to forecasting, alongside time-series classification, regression and clustering in the same library. Its particular strength is validation: temporal train/test splitting, rolling-window cross-validation, and pipeline composition with detrending and deseasonalizing transformers — for teams building rigorous, leakage-free evaluation rather than fitting a model ad hoc.
• Direct architectural control over deep learning. GluonTS provides reference implementations of neural forecasting architectures (DeepAR, Temporal Fusion Transformer and others) with the data-loading and probabilistic-output plumbing to train them, aimed at researchers and engineers who want to work at the architecture level rather than have a tool choose for them — it sits a level below AutoGluon in abstraction, and AutoGluon can automate selection on top of libraries like it.

Statistical vs. machine-learning vs. deep-learning models

This category spans three model families, and most libraries mix them rather than committing to one:

• Statistical models (ARIMA, exponential smoothing, Theta) are fast, interpretable, and strong baselines — the core of Nixtla's StatsForecast and available inside Darts and sktime.
• Gradient-boosting models (LightGBM, XGBoost) often outperform both on business time series with good features, and appear inside Darts, Nixtla's MLForecast and AutoGluon's ensemble pool.
• Deep-learning models (N-BEATS, Temporal Fusion Transformer, DeepAR) can capture complex patterns across many related series but need more data, more compute, and more expertise — the focus of PyTorch Forecasting and GluonTS, and available inside Darts, AutoGluon and Nixtla's NeuralForecast.

Don't assume deep learning wins by default. On a single, well-behaved business series, a well-tuned statistical model or Prophet often matches or beats a neural network with far less engineering effort. Deep learning earns its complexity when you have many related series with shared patterns to learn from, or genuinely large-scale, irregular data.

Deployment and compute

All seven are self-hosted by default — code you run in your own infrastructure — except Nixtla's TimeGPT and AutoGluon's optional SageMaker integration, both of which are cloud services. Deep-learning libraries (PyTorch Forecasting, GluonTS, and Darts or AutoGluon when using their neural models) benefit substantially from GPU access for training; the statistical and gradient-boosting paths (Prophet, sktime, Nixtla's StatsForecast) do not need one.

Pricing: free, except one thing

Every library here is free and open source with no commercial tier. The one exception is Nixtla's TimeGPT, a paid, quote-based hosted service layered on top of Nixtla's otherwise-free open-source libraries — everything else in this category costs only engineering time and, for deep-learning models, compute.

A shortlist by situation

• If you need a fast, interpretable forecast on one series with holidays and clear seasonality, look at Prophet.
• If you're forecasting thousands to millions of series at once, look at Nixtla's StatsForecast or MLForecast, or its hosted TimeGPT if you don't want to train anything.
• If you want a strong automated baseline with no model-selection expertise, look at AutoGluon.
• If you want to benchmark many model families side by side through one consistent API, look at Darts.
• If you need rigorous, leakage-free evaluation using scikit-learn-style pipelines, look at sktime.
• If you're forecasting a large panel of related series (many stores, SKUs or sensors) and want deep learning with covariates, look at PyTorch Forecasting.
• If you want direct control over neural forecasting architecture as a researcher or ML engineer, look at GluonTS.

Questions to ask, or check, before committing

• How many series do we actually need to forecast, and does this library's design assume one series or many?
• Do we have the GPU access and ML expertise a deep-learning approach requires, or would a statistical baseline get us most of the way there for less effort?
• Does the library support backtesting and cross-validation in a way that avoids leaking future data into training?
• If we start with an automated tool like AutoGluon, can we drop down to a more controllable library later without starting over?
• For Nixtla's TimeGPT specifically: what does it cost at our forecast volume, since it's the one paid product in an otherwise free category?

Common mistakes

• Reaching for a deep-learning library by default when a well-tuned statistical model or Prophet would forecast a single series just as well, faster.
• Treating an automated tool's output as final without backtesting it against a held-out period.
• Ignoring seasonality structure in the data and expecting any model to compensate for features that were never given to it.
• Underestimating the GPU and engineering cost of a deep-learning approach because the library itself is free.

For direct head-to-heads, see Darts vs sktime and Nixtla vs Prophet. For every tool in this category, browse the full directory.