Deep Learning Architectures for Time-Series Forecasting and Classification
Designing neural architectures, from transformers to simple linear and MLP models, for forecasting and classifying time series, measured against statistical baselines.
Tier Grades
Tier A: Frontier AI labs
feasible
Google, Amazon and universities publish forecasting architectures.
Tier B: Indonesian AI labs
feasible
Models of this kind train on single GPUs; the anchor papers come from 4-author teams.
Tier C: Institution units such as ITB
feasible
PatchTST and DLinear each came from 4-author teams.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
feasible
Demand and energy forecasting proofs of value use these models with statistical baselines.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeNo compute figures; PatchTST cuts attention compute and memory quadratically by patching; LTSF-Linear is a one-layer linear model.
- DataNine real-life datasets on which one-layer LTSF-Linear beats transformer forecasters (DLinear paper); PatchTST adds masked self-supervised pretraining.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 4 authors on PatchTST; 4 on the DLinear paper.
- Time horizonMonths to a year.
- Shelf-life signalsSimple linear baselines beating transformers (DLinear) and pretrained foundation forecasters taking over, Classical baselines (ETS, Theta, seasonal naive) kept as the yardstick
- Pipeline roleTier A and universities publish architectures; Tier C benchmarks them on Indonesian series; Tier D builds forecasting proofs of value.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
Observed actors
Linked models
PatchTSTtime seriesexemplar
DLinear and NLineartime seriesexemplar
Temporal Fusion Transformer (TFT)time seriesexemplar
N-BEATStime seriesexemplar
Granite TimeSeries TTM r3time seriesproduced by
TimeMixertime seriesproduced byGranite TinyTimeMixer (TTM) r2time seriesproduced by
iTransformertime seriesproduced by
TiDEtime seriesproduced byTSMixertime seriesproduced byTimesNettime seriesproduced by
FEDformertime seriesproduced byN-HiTStime seriesproduced by
MSTLtime seriesused byAutoformertime seriesproduced byInformertime seriesproduced by
ROCKETtime seriesproduced by
InceptionTimetime seriesproduced by
TCN (Temporal Convolutional Network)time seriesproduced by
MQ-RNN and MQ-CNNtime seriesproduced byDeepARtime seriesproduced byLSTNettime seriesproduced byAuto_Arimatime seriesused byGranite TinyTimeMixer (TTM) r2.1time seriesproduced byLinks
R&D TopicDeep Learning Architectures for Time-Series Forecasting and Classification
ModelsDLinear and NLinearThe Chinese University of Hong Kong, time seriesN-BEATSElement AI, time seriesPatchTSTPrinceton University, time seriesTemporal Fusion Transformer (TFT)University of Oxford, time series+27 more
DLinear and NLinearThe Chinese University of Hong Kong, time seriesN-BEATSElement AI, time seriesPatchTSTPrinceton University, time seriesTemporal Fusion Transformer (TFT)University of Oxford, time series+27 moreUse CasesIndustrial Forecasting (Load, Throughput, Demand)IndustrialDemand ForecastingRetail and Logistics
Who Runs Itnone recorded
Models 31
Granite TinyTimeMixer (TTM) r2IBM, time seriesproduced byGranite TinyTimeMixer (TTM) r2.1IBM, time seriesproduced byGranite TimeSeries TTM r3IBM, time seriesproduced byTemporal Fusion Transformer (TFT)University of Oxford, Google, time seriesexemplarDLinear and NLinearThe Chinese University of Hong Kong, IDEA Research (International Digital Economy Academy), time seriesexemplar
Intel time-series-forecasting-electricity-0001Intel, time seriesproduced byMQ-RNN and MQ-CNNAmazon, time seriesproduced byMSTLUniversity of Melbourne, time seriesused byN-BEATSElement AI, Mila, time seriesexemplarPatchTSTPrinceton University, IBM, time seriesexemplarTCN (Temporal Convolutional Network)Carnegie Mellon University, Intel, time seriesproduced byTimeMixerAnt Group, Tsinghua University, time seriesproduced by+19
TimesNetTsinghua University, time seriesproduced byAutoformerTsinghua University, time seriesproduced byCroston methodtime seriesused byDeepARAmazon, time seriesproduced byETS (exponential smoothing)time seriesused byFEDformerAlibaba Group, time seriesproduced byInceptionTimeUniversite de Haute-Alsace, time seriesproduced byiTransformerTsinghua University, time seriesproduced byLSTNetCarnegie Mellon University, time seriesproduced by
Nixtla statsforecastNixtla, time seriesused by
ProphetMeta, time seriesused byROCKETMonash University, time seriesproduced bySeasonal_Naivetime seriesused byTheta methodtime seriesused byTiDEGoogle, time seriesproduced byTSMixerGoogle, time seriesproduced byAuto_Arimatime seriesused byInformertime seriesproduced byN-HiTStime seriesproduced byPapers
- DeepAR: Probabilistic Forecasting with Autoregressive Recurrent NetworksDavid Salinas et al., arXiv preprint, 2017 founding
- iTransformer: Inverted Transformers Are Effective for Time Series ForecastingYong Liu et al., arXiv preprint, 2023 state of the art
- Are Transformers Effective for Time Series Forecasting?Ailing Zeng et al., arXiv preprint, 2022 state of the art
- A Time Series is Worth 64 Words: Long-term Forecasting with TransformersYuqi Nie et al., ICLR 2023 (per arXiv comment), 2022 state of the art
- Transformers in Time Series: A SurveyQingsong Wen et al., In the 32nd International Joint Conference on Artificial Intelligence (IJCAI 202, 2022 survey
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- arxiv.org/abs/2211.14730 https://arxiv.org/abs/2211.14730
- arxiv.org/abs/2205.13504 https://arxiv.org/abs/2205.13504
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Deep Learning Architectures for Time-Series Forecasting and Classification", https://aurion.semai.id/topics/deep-learning-architectures-for-time-series/, records of October 2026.