Information Hub

Articles and updates from SoloQuant Academy.

Tag: paper

pythonpapertime-seriestransformerssurveyafter-work

Introducing Transformers in Time Series: A Survey for after-work quants

Wen, Zhou, Zhang, Chen, Ma, Yan and Sun (IJCAI 2023; arXiv:2202.07125) survey Transformers for time series. Taxonomy: module tweaks (positional encodings, sparse/low-rank attention) vs architecture (hierarchical). Applications: forecasting, anomaly, classification. Their ETTm2 checks: longer inputs often hurt; 3–6 layers beat 48; seasonal-trend decomposition can cut MSE 50–80%. Map before you pick Informer vs PatchTST. Plain-language intro. Not investment advice.

pythonpapertime-seriesitransformerforecastingafter-work

Introducing iTransformer: Inverted Transformers Are Effective for Time Series Forecasting for after-work quants

Liu, Hu, Zhang, Wu, Wang, Ma and Long (ICLR 2024; arXiv:2310.06625) invert the Transformer: each variate’s full look-back is one token. Attention mixes assets; FFN (shared MLP) encodes that series; LayerNorm is per variate. No new attention gadget. Strong on high-dimensional ECL/Traffic/PEMS; inverting also lifts vanilla/Informer/Reformer. thuml/iTransformer. Plain-language intro. Not investment advice.

pythonpapertime-seriestimesnetforecastingafter-work

Introducing TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis for after-work quants

Wu, Hu, Liu, Zhou, Wang and Long (ICLR 2023; arXiv:2210.02186) argue 1D series hide intraperiod vs interperiod variation. FFT picks top-k periods, reshape each into a 2D tensor (columns = inside a period, rows = same phase across periods), run a shared Inception block, softmax-amplitude mix. SOTA-ish across forecasting, imputation, UEA classification, anomaly detection. thuml/TimesNet. Plain-language intro. Not investment advice.

pythonpapertime-seriestftforecastingafter-work

Introducing Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting for after-work quants

Lim, Arik, Loeff and Pfister (IJF 2021; arXiv:1912.09363) build TFT for multi-horizon forecasts that mix static covariates, past-observed series, and known-future inputs (calendar, promotions). GRNs + GLU skip unused parts; variable selection names which features matter; LSTM does local processing; interpretable multi-head attention does long range; quantiles 0.1/0.5/0.9. Electricity, Traffic, Favorita, volatility. Plain-language intro. Not investment advice.

pythonpapertime-seriesnhitsforecastingafter-work

Introducing N-HiTS: Neural Hierarchical Interpolation for Time Series Forecasting for after-work quants

Challu, Olivares, Oreshkin, Garza, Mergenthaler-Canseco and Dubrawski (AAAI 2023; arXiv:2201.12886) keep N-BEATS residual MLP stacks, then add MaxPool multi-rate sampling and hierarchical interpolation so each block owns a frequency band. Univariate, shared weights. About 16% MSE / 14% MAE vs Transformer LTSF baselines, about 50 times faster. NeuralForecast. Plain-language intro. Not investment advice.

pythonpapertime-seriestransformersurveyafter-work

Introducing Transformers in Time Series: A Survey for after-work quants

Wen, Zhou, Zhang, Chen, Ma, Yan and Sun (IJCAI 2023; arXiv:2202.07125) map TS Transformers by module (position, attention, hierarchy) and by task (forecast, anomaly, classify). Complexity table, ETTm2 length/depth/decomposition probes, GitHub living list. Plain-language intro. Not investment advice.

pythonpapertime-seriestransformersurveyafter-work

Introducing Transformers in Time Series: A Survey for after-work quants

Wen, Zhou, Zhang, Chen, Ma, Yan and Sun (IJCAI 2023; arXiv:2202.07125) survey Transformers for time series. Taxonomy: module tweaks (positional encodings, sparse/low-rank attention) vs architecture (hierarchical). Applications: forecasting, anomaly, classification. Their ETTm2 checks: longer inputs often hurt; 3–6 layers beat 48; seasonal-trend decomposition can cut MSE 50–80%. Map before you pick Informer vs PatchTST. Plain-language intro. Not investment advice.

pythonpapertime-seriesitransformerforecastingafter-work

Introducing iTransformer: Inverted Transformers Are Effective for Time Series Forecasting for after-work quants

Liu, Hu, Zhang, Wu, Wang, Ma and Long (ICLR 2024; arXiv:2310.06625) invert the Transformer: each variate’s full look-back is one token. Attention mixes assets; FFN (shared MLP) encodes that series; LayerNorm is per variate. No new attention gadget. Strong on high-dimensional ECL/Traffic/PEMS; inverting also lifts vanilla/Informer/Reformer. thuml/iTransformer. Plain-language intro. Not investment advice.

pythonpapertime-seriestimesnetforecastingafter-work

Introducing TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis for after-work quants

Wu, Hu, Liu, Zhou, Wang and Long (ICLR 2023; arXiv:2210.02186) argue 1D series hide intraperiod vs interperiod variation. FFT picks top-k periods, reshape each into a 2D tensor (columns = inside a period, rows = same phase across periods), run a shared Inception block, softmax-amplitude mix. SOTA-ish across forecasting, imputation, UEA classification, anomaly detection. thuml/TimesNet. Plain-language intro. Not investment advice.

pythonpapertime-seriestimesnetforecastingafter-work

Introducing TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis for after-work quants

Wu, Hu, Liu, Zhou, Wang and Long (ICLR 2023; arXiv:2210.02186) argue 1D series hide intraperiod vs interperiod variation. FFT picks top-k periods, reshape each into a 2D tensor (columns = inside a period, rows = same phase across periods), run a shared Inception block, softmax-amplitude mix. SOTA-ish across forecasting, imputation, UEA classification, anomaly detection. thuml/TimesNet. Plain-language intro. Not investment advice.

pythonpapertime-seriestftforecastingafter-work

Introducing Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting for after-work quants

Lim, Arik, Loeff and Pfister (IJF 2021; arXiv:1912.09363) build TFT for multi-horizon forecasts that mix static covariates, past-observed series, and known-future inputs (calendar, promotions). GRNs + GLU skip unused parts; variable selection names which features matter; LSTM does local processing; interpretable multi-head attention does long range; quantiles 0.1/0.5/0.9. Electricity, Traffic, Favorita, volatility. Plain-language intro. Not investment advice.

pythonpapertime-seriesnhitsforecastingafter-work

Introducing N-HiTS: Neural Hierarchical Interpolation for Time Series Forecasting for after-work quants

Challu, Olivares, Oreshkin, Garza, Mergenthaler-Canseco and Dubrawski (AAAI 2023; arXiv:2201.12886) keep N-BEATS residual MLP stacks, then add MaxPool multi-rate sampling and hierarchical interpolation so each block owns a frequency band. Univariate, shared weights. About 16% MSE / 14% MAE vs Transformer LTSF baselines, about 50 times faster. NeuralForecast. Plain-language intro. Not investment advice.