Figure 11 — Downstream tasks and models

Vector PDF

File: fig11_tasks_models.pdf

Suggested caption. The 14 downstream tasks (U2.1–U2.9) and the models trained on each. All tasks use the same 11 features. Labels are functions of the series, with thresholds frozen on real train. Task U2.1 (green) is the task of the pre-specified case-study cell.

Tasks

ID	Task(s)	Kind	Target
U2.1	clf_up_h24_q90	binary	change over 24 h above the 0.90 train quantile
U2.2	clf_up_h{1,6,168}_q90	binary	same, other horizons
U2.3	clf_up_h24_q{95,99}	binary	rare events
U2.4	clf_down_h24_q10	binary	change below the 0.10 quantile (opposite tail)
U2.5	mc3_h24	3 classes	down / neutral / up
U2.6	reg_h{1,24,168}	regression	value x_{t+h}
U2.7	qreg_h24	quantile regression	quantiles 0.1 / 0.5 / 0.9
U2.8	ms_h24	multistep	vector x_{t+1..t+24}
U2.9	vol_h24	volatility	std of the next 24 hourly changes

Models

• linear: logistic/softmax regression (L-BFGS, balanced weights, C = 1), ridge (α = 1, exact solve), linear quantile regression (smoothed pinball). PyTorch.
• mlp: 64 → 32 ReLU network, Adam lr 1e-3, weight decay 1e-4, batch 512. In min/mid it runs 30 epochs on U2.1 only; in max, up to 200 epochs with early stopping on all tasks.
• gbdt: XGBoost. min/mid: 100 trees, depth 4, 64 bins, CPU, vector leaves for the multistep task. max: 200 trees, depth 6, 256 bins, one ensemble per step.

Notes for the text

• Every (task, model) pair is fitted in both regimes (S and R) and evaluated on both val and OOT.
• Every task enters the ranking and the research questions with equal weight, grouped into classification and regression; U2.1 only marks the case-study cell (CS7), U2.2–U2.9 are case study CS8.
• Source: marathonts/bench/tasks.py, marathonts/bench/models.py, MODELS.md, METRICS.md (U2–U4).