SoloQuant Flagship Program Early Bird · $420 USD (Save $379) · Sep 1–Dec 1, 2026 · Launch Jan 1, 2027 →

SoloQuant Academy

Short evening lessons for people who work. Quant trading, Python, data, systems, architecture.

Information Hub Student Portal Privacy Policy Terms & Conditions

Academy notes

Information Hub

Articles and updates from SoloQuant Academy.

Tag: time-series

pythonpapertime-seriesfedformertransformerafter-work

Introducing FEDformer: Frequency Enhanced Decomposed Transformer for Long-term Series Forecasting for after-work quants

Zhou, Ma, Wen, Wang, Sun and Jin (Alibaba, ICML 2022; arXiv:2201.12740) mix seasonal-trend decomposition with frequency-domain Transformer blocks. Randomly keep a fixed number of Fourier (or Wavelet) modes so attention is linear in length. vs Autoformer: about 14.8% / 22.6% lower MSE on multivariate / univariate LTSF. Plain-language intro. Not investment advice.

python

paper

time-series

moment

foundation-model

after-work

Introducing MOMENT: A Family of Open Time-series Foundation Models for after-work quants

Goswami, Szafer, Choudhry, Cai, Li and Dubrawski (CMU Auton Lab, ICML 2024; arXiv:2402.03885) release open TS foundation models (40M / 125M / 385M). Masked patch reconstruction on the Time Series Pile; one encoder for forecast, classify, impute, anomaly. Hugging Face AutonLab/MOMENT-1-large. Plain-language intro. Not investment advice.

pythonpapertime-serieslstmattentionafter-work

Introducing EA-LSTM: Evolutionary Attention-based LSTM for Time Series Prediction for after-work quants

Li, Zhu, Kong, Han and Zhao (2018; arXiv:1811.03760) put an evolutionary attention layer in front of an LSTM. Attention weights over the look-back window are evolved with competitive random search (genetic-style select/recombine/mutate), not backprop, to avoid local traps. Beijing PM2.5, SML2010, MSR Action3D. Plain-language intro. Not investment advice.

pythonpapertime-seriesinformertransformerafter-work

Introducing Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting for after-work quants

Zhou, Zhang, Peng, Zhang, Li, Xiong and Zhang (AAAI 2021; arXiv:2012.07436) fix three Transformer LSTF bottlenecks: quadratic attention (ProbSparse O(L log L)), stacked-layer memory (self-attention distilling), and slow step-by-step decode (one-shot generative decoder). ETT oil temperature, ECL, Weather. Plain-language intro. Not investment advice.

pythonpapertime-seriesfedformertransformerafter-work

Introducing FEDformer: Frequency Enhanced Decomposed Transformer for Long-term Series Forecasting for after-work quants

pythonpapertime-serieslstmattentionafter-work

Introducing EA-LSTM: Evolutionary Attention-based LSTM for Time Series Prediction for after-work quants

pythonpapertime-serieschronosfoundation-modelafter-work

Introducing Chronos: Learning the Language of Time Series for after-work quants

Ansari, Stella, Turkmen et al. (AWS AI Labs, 2024; arXiv:2403.07815) treat forecasting as language modeling: mean-scale, uniform bins (~4096), T5 20M–710M, cross-entropy. Pretrain on public + TSMixup + KernelSynth GPs. Strong in-domain and competitive zero-shot on 42 datasets. Code: amazon-science/chronos-forecasting. Plain-language intro. Not investment advice.

pythonpapertime-seriesdeeparprobabilisticafter-work

Introducing DeepAR: Probabilistic Forecasting with Autoregressive Recurrent Networks for after-work quants

Salinas, Flunkert and Gasthaus (Amazon, 2017; IJF 2020; arXiv:1704.04110) train one autoregressive LSTM on thousands of related series. The net emits likelihood parameters (Gaussian or negative binomial); ancestral sampling yields joint predictive paths and consistent quantiles. Scale handling + velocity-weighted sampling for power-law magnitudes. Plain-language intro. Not investment advice.

pythonpapertime-seriescnntransformerafter-work

Introducing Financial Time Series Forecasting using CNN and Transformer for after-work quants

Zeng, Kaur, Siddagangappa, Rahimi, Balch and Veloso (J.P. Morgan AI Research, 2023; arXiv:2304.04912) propose CTTS: min-max scale, 1D CNN (k=16, stride 8) tokens, 4-layer Transformer, MLP softmax over up/down/flat. 1-min S&P 500 2019; 80 steps → sign at 81. 2-class 56.7% vs DeepAR 51.1%. Plain-language intro. Not investment advice.

pythonpapertime-seriesautoformertransformerafter-work

Introducing Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting for after-work quants

Wu, Xu, Wang and Long (NeurIPS 2021) put moving-average decomposition inside the Transformer and replace point-wise attention with FFT Auto-Correlation (O(L log L)). ~38% average MSE drop vs Informer/LogTrans on energy, traffic, FX, weather, ILI. Plain-language intro. Not investment advice.

pythonpapertime-seriesdlineartransformerafter-work

Introducing Are Transformers Effective for Time Series Forecasting? (DLinear / LTSF-Linear) for after-work quants

Zeng, Chen, Zhang and Xu (2022) ask whether Transformers help long-term time-series forecasting. One-layer LTSF-Linear (DLinear / NLinear) beats Informer, Autoformer, FEDformer on nine benchmarks, often by 20–50% MSE. Plain-language intro. Not investment advice.

pythonpapertime-seriestransformerafter-work

Introducing A Time Series is Worth 64 Words: Long-term Forecasting with Transformers (PatchTST) for after-work quants

Nie, Nguyen, Sinthong and Kalagnanam (2022) treat each univariate channel as patches (ViT-style) sharing one Transformer. Beats FEDformer/Autoformer/Informer and often DLinear on ETT/Traffic/Electricity/Weather. Plain-language intro. Not investment advice.

← Previous

Page 3 of 4