Théo Charron - Data Analyst

A 3-state HMM denoises multi-horizon OLS slope. Stable soft-fade structure, useful mid-episode hazard — no entry gate or net EV.

In brief. I filtered OLS slope t-stat through a 3-state HMM (down / neutral / up) on horizons . Soft-fade episodes (short doubt vs long thesis) show stable structure: CIF R/E/B ≈ 28/68/4% IS≈OOS. Mid-episode risk is readable (AUC E ≈ 0.82), not at onset (AUC ≈ 0.52). Net EV ≈ 0 after costs.

Why I ran this study

Raw slope is noisy; MAP alone is unstable bar-to-bar. Idea: causal soft slopes as a regime layer, then test whether a “fade” pattern (s16 turns neutral while s128 stays on thesis) carries tradable information.

Classic trap: confusing a high conditional realignment rate with a favourable onset probability.

Setup (summary)

Parameter	Value
Asset / barrier	EURUSD · TIMB
HMM fit	60d 2023 → freeze
Eval	IS post-calib 2023 + OOS 2024
HMM	C2, K=3, sticky E[D]=40

Soft-fade onset (e.g. long): drops from +1 to 0 while .

Absorbing outcomes: Resolve (R), Escalate (E), Break (B).

Key results

CIF issues soft-fade — Resolve / Escalate / Break (IS vs OOS 2024)

Gate	Verdict	Detail
0 Setup	PASS	OOS freeze corr 0.955–0.996
1 Existence	PASS	Stable CIF IS/OOS
2 Detectability	FAIL	AUC E onset 0.522
3 Predictability	PARTIAL	AUC E mid 0.816
4 Tradability	FAIL	mean pips @0.2 RT ≈ −0.21 OOS

Gate 2 vs 3 — AUC Escalate (OOS) : onset ≈ hasard, mid-épisode ≈ 0,82

PnL by outcome: resolve ~+4.3 pips, escalate ~−1.8 — with ~68% escalate, the book dies.

Takeaways

Question	Answer
Is the denoiser sound?	Yes — fast (~23 s/yr), stable freeze
Entry gate at onset?	No
Dynamic mid-episode risk?	Yes — person-period
Standalone trade?	No

“~90% realignment” is conditional on R∪B — not P(winning onset).

Recommended use: multi-scale softs as regime / hazard features — not an isolated entry signal.

HMM slope denoising — regime yes, entry no

Why I ran this study

Setup (summary)

Key results

Takeaways

Links