Free tools Windows power users keep installed
One-click scans. No signup required.
No single deep-learning architecture is best for every univariate time series. The right choice depends on the series, forecast horizon, evaluation method, available training data and compute, and whether you need a point forecast or a probability distribution. For a fair decision, compare candidate models on the same chronological forecast task and include simple statistical or naïve baselines.
What univariate forecasting means—and what it does not
Univariate forecasting predicts future values of one target series from that series’ own history. A target-only model does not use external inputs such as weather, prices, or calendar features. Some forecasting studies include covariates, so their results may not transfer directly to a target-only setup.
The forecast task also matters. In one-step forecasting, the model predicts the next value. In multi-horizon forecasting, it predicts several future values, either by producing them directly or by repeatedly feeding predictions back into the model. A model can perform differently across these settings, so a comparison must fix the horizon and prediction strategy.
Finally, a point forecast gives a single predicted value for each time step. A probabilistic forecast represents uncertainty, for example through a predictive distribution or intervals. Not every model or implementation provides the latter, and point-forecast accuracy alone does not show whether uncertainty estimates are useful.
Recommended Free Tools
#1 Best Overall
Which model families are used?
Deep-learning approaches differ in how they represent the sequence and produce future values. The family name alone does not determine suitability: the specific architecture, training setup, data, and forecast task all matter.
Feed-forward networks and MLPs
Feed-forward models map a fixed window of past observations to future values without recurrently processing each time step. N-BEATS is a prominent example framed as a univariate point-forecasting model; its original paper describes forecasting through stacks of basis-expansion blocks. N-HiTS is another MLP-based approach included in the NeurIPS 2023 forecasting comparison. This family is a practical candidate when you want a direct output for a defined forecast horizon.
Recurrent networks
RNNs process observations in sequence, updating an internal state as they move through the history. LSTMs are a recurrent variant designed to handle longer-range dependencies. They remain useful conventional neural baselines in forecasting surveys, but their presence in the literature does not establish that they will outperform other choices on a particular series.
CNNs and temporal convolutional networks
Convolutional models apply filters across time to detect local patterns. Temporal convolutional networks (TCNs) use temporal convolutions and receptive fields to incorporate information from a wider history. They offer a different way to model temporal structure from recurrent networks or attention-based models.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Transformers and attention-based models
Transformers use attention mechanisms to relate parts of a sequence. PatchTST, for example, segments the input history into patches before applying a Transformer-based approach. Other Transformer variants appear in the broader time-series literature. Compare a specific implementation under your own task and training protocol rather than assuming that a newer or more prominent architecture will win.
Other approaches in broader surveys
Recent surveys also cover graph neural networks (GNNs), diffusion models, and large-language-model approaches alongside MLPs, RNNs, CNNs, and Transformers. Their inclusion in a survey is not evidence that they are appropriate for a target-only univariate task. Before treating one as a direct alternative, check whether the method is designed and evaluated for that setting.
Rank #4
- Used Book in Good Condition
How to compare models fairly
Forecasting scores only mean something in the context of the data, horizon, split, and metric used to produce them. Set up the comparison so that each candidate faces the same task.
- Define the inputs and output. Record whether the task is target-only or uses covariates, whether it is one-step or multi-horizon, and whether you require point predictions or probabilistic forecasts.
- Fix chronological train, validation, and test windows. Do not let future observations leak into training or model selection. Use multiple rolling forecast origins when appropriate, applying the same windows to each candidate.
- Choose metrics that match the decision. Scale-dependent errors are useful when errors in the original units matter; scale-independent metrics help compare series on different scales. If forecast distributions matter, use probabilistic scoring rather than relying only on point-error metrics.
- Keep simple baselines in the comparison. Include naïve or statistical forecasts even if the main question is which deep-learning model to use. A complex model is only useful if it improves on an appropriate baseline for the task.
- Track operational costs and data needs. Alongside forecast quality, measure training time, inference latency, memory use, and the amount of training data required. Check whether the implementation supports the forecast intervals or distributions your application needs.
Do not select a winner from scores measured on different splits, horizons, or metrics. If you change any of those conditions, treat it as a different forecasting task and compare candidates again.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Best Value
What benchmark results can—and cannot—tell you
The M4 competition is useful historical context, not a guarantee about performance on a new series. The Royal Society survey (2021) describes M4 as involving 100,000 time series and 61 forecasting methods. That scale makes it a substantial comparison, but a result from a competition still belongs to its datasets, forecast setup, and evaluation protocol.
The NeurIPS 2023 paper’s Section 5.1 reports weighted-average sMAPE, MASE, and OWA values for multiple models in a univariate M4 results table. Those rankings should be read with that table’s setup and metrics attached; they are not a universal ordering of architectures or a prediction of how the same models will perform on your data. The reported metrics are also not automatically the right ones for every application.
A practical starting point
For a target-only task, start by specifying the history window, forecast horizon, and whether you need a point or probabilistic output. Then compare a small set of candidates from relevant families—such as an MLP model, a recurrent model, a convolutional model, or a Transformer—against naïve or statistical baselines on identical chronological windows. Choose using both forecast quality and operational requirements, not architecture labels or a result from a different evaluation.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
