Figure 5 — Temporal split and the role of each part

Vector PDF

File: fig05_temporal_split.pdf

Suggested caption. Temporal split (profile min, drawn to scale). Hatched bars are embargo gaps of 168 h, wide enough to cover every label horizon h ≤ 168 and the longest feature lag (168). Inside train, the last 10% is held out to stop generator training early. Every statistic the methods may use comes from train. Val is the primary fidelity reference; train and OOT are also references. Val and OOT are the utility test sets.

How to read it

• Top row: train (50%) | embargo | val (25%) | embargo | OOT (25%). In profile max the shares are 60/20/20 over the full series.
• Bottom row: train broken down for the generators: core windows | embargo (168 rows) | holdout. Core and holdout windows do not overlap.
• Labels above list what each part is used for.

Design rationale (for the paper)

• Fidelity is measured against three real references: val (primary, no suffix), train (@train) and OOT (@oot). Train shows how well a method reproduces the data it was fitted on — methods fitted on train, and methods that take reference statistics from train, use this data during fitting; val measures generalisation and OOT generalisation under drift. Two of the four rankings measure fidelity against val and OOT (holdout_clf, holdout_reg), two against train (train_clf, train_reg).
• Real-versus-real reference. Pseudo-method real uses the OOT series as a "synthetic" series (RQ0); it is a reference, not an upper bound, since OOT is shifted in time. OOT is shorter than K·T, so each method is also scored on its first len(OOT) points (@len) to compare at equal length.
• Utility uses the same real val and OOT rows for TSTR and TRTR. OOT checks robustness to a shift in time (RQ5).