Add DLinear-t results

[sereiwathnaa]

Model. DLinear-t is a direct distributional forecaster trained from scratch per configuration: a DLinear conditional-mean backbone (per-window instance normalisation) and a trailing-variance MLP feed a one-step Student-t output whose tail is fixed at ν = 5.001. One forward pass yields all horizon-wise parameters; forecast samples are independent draws (no diffusion, no autoregression).

Protocol. One model per configuration, trained on the benchmark's validation split (train + one validation horizon, as the deep-learning baselines), univariate (to_univariate=True for multivariate datasets), 100 samples per forecast, evaluated with gluonts evaluate_model (axis=None, mask_invalid_label=True, allow_nan_forecast=False, benchmark seasonality). The last H observations of every training series are held out for early stopping. No test data is used for training or model selection. Seed 1; 3000 / 5000 / 6000 Adam steps for short / medium / long terms with warm-up + cosine decay, patience 8 on the held-out NLL.

Results (97 configurations). CRPS/Seasonal-Naive 0.632, MASE/Seasonal-Naive 0.869 (geometric means).

Files. results/DLinear-t/all_results.csv, results/DLinear-t/config.json, and the replication notebook notebooks/dlinear_t.ipynb (self-contained: gift-eval + gluonts + torch; ~45 GPU-minutes for all 97 configurations).

• model_type: deep-learning, float32
• testdata_leakage: No
• replication_code_available: Yes
• org: Big Data CBNU
• contact: sereiwathnaros@chungbuk.ac.kr

[salesforce-cla]

Thanks for the contribution! Before we can merge this, we need @sereiwathnaa to sign the Salesforce Inc. Contributor License Agreement.

[cuthalionn]

Hi @sereiwathnaa , thanks for the submission. I was reproduced the results, the model results are in the leaderboard now.