Notes | Vincent Wang Research

Frictions all the way down: the implementation gap, layer by layer 2026-06-19 Reporting one top-line Sharpe is a claim about the cleanest possible layer. Measuring the same strategy at the instrument and integer-contract layers — and naming which step costs the most — is more honest, and it changes where you spend your effort.
Arithmetic can be legislated; capital allocation cannot 2026-06-19 Pre-registered thresholds are a commitment device for the things that should be mechanical. The failure mode is overreach — the moment a rule starts deciding how much capital is at risk, it has left the domain where pre-registration protects you.
Look-ahead prevention as a unit test, not a promise 2026-06-19 "We were careful" is not a control. Perturb the future and assert the present does not move — a direct, mechanical check for look-ahead that keeps working as the codebase changes.
Selection bias in 0DTE breakout backtests: how full-population validation overturned four positive drafts Four drafts of the Donchian 0DTE paper reported +$80K, +$50K, and +$32K per year. Full-population tick validation reported −$107K. The gap was two compounding biases — and the larger one was a selection filter that looked like engineering hygiene.
Black-Scholes with VIX1D under-prices ATM 0DTE entry premium by roughly 85% Across thousands of validated trades, real entry mid / BSM entry premium has a median of 1.85. The error is stable, asymmetric between entry and exit, and large enough to manufacture phantom edge on a signal that has none.
Dual-clock backtesting: minute-level signals with a 1-second risk clock Signals fire on closed minute bars; stops, targets, and trailing logic are evaluated on 1-second closes with a strict per-second event order. Validation showed minute-only risk evaluation distorts tight-stop strategies by −55% while leaving wide-stop profiles almost untouched.
Cross-day rolling Donchian channels for valid day-one signals With a 30-bar 5-minute lookback computed within-session, no signal can exist for the first 2.5 hours of each day. Rolling the channel across session boundaries restores the open — at the cost of overnight-gap distortion that needs its own policy.
Option-proxy design: BSM, delta proxy, and VIX-scaled IV versus real SPXW quotes The research ladder prices option exposure three increasingly honest ways — delta proxy, BSM with a VIX-scaled IV, BSM with real/fitted IV — and reserves the tradability verdict for real quote-chain replay. Each proxy has a defined role and a defined ceiling.
Walk-forward, OOS, Deflated Sharpe, and PBO: overfitting controls that matter The controls that did the work here are walk-forward gates with in-window selection, bootstrap confidence intervals, dominance checks, and parameter-stability evidence. Deflated Sharpe and PBO are specified as acceptance criteria but deliberately deferred — and saying so plainly is part of the methodology.
Risk invariants and execution safety in live 0DTE systems Safety outranks correctness, which outranks profit. The live 0DTE systems encode that as hard invariants — single position, long-only, daily hard stops, forced EOD flatten — enforced by state machines and schema-level tests rather than comments, with production incidents converted into permanent regression tests.
Why next-open execution can still be optimistic Signal at close, execute at next open is the standard honest daily convention — and it still assumes a fill nobody guarantees. Lag-grid evidence from intraday work shows edges decaying within seconds, and spread measurements show exits costing four times entries.
The window trap: unequal histories make rankings lie 2026-06-25 Full-available rankings mix long and short histories, so they reward an accident of inception. Re-rank on a single equal window before believing any "best performer" claim — the leader can flip.
A monthly crisis window can hide the crash 2026-06-25 Crisis performance measured on monthly returns can post a gain while the asset fell 38% peak-to-trough inside the very same window. Report intra-window daily drawdown to measure a crash faithfully.
'Defensive' is a claim that needs a significance test 2026-06-25 A point-estimate ranking of "defensive" assets is not evidence of a defensive edge. Attach n, t, and p before believing it — here the highest mean was insignificant, and the reliable hedge had the smaller mean.
73 corrections or 339? The counting convention decides 2026-07-10 Anchor drawdowns at all-time highs and a 98-year history has 73 episodes of 5%+; anchor them at local peaks and it has 339. Neither is wrong — they answer different questions. Never quote a drawdown count without its convention.
The median bear recovers in 764 days — the error bars span 666 to 2310 2026-07-10 Recovery-time "averages" come from a dozen observations, so their uncertainty is enormous — the 95% CI on the median 20%+ recovery is [666, 2310] days. Meanwhile half the loss is typically back in 220 days. Quote the interval and the milestones, not just the point estimate.
In income-scaled DCA, the window picks the winner 2026-07-10 Money-weighted returns inherit the window trap — with growing contributions, late-cycle dollars dominate, and own-inception rankings mix unequal histories. Compare DCA outcomes only on a common window where every asset receives the same cash-flow schedule.
The universe was chosen after the race 2026-07-10 Any study whose asset list contains known ex-post winners carries hindsight selection and survivorship bias by construction. Read counterfactual DCA results as expectation-setting against a broad benchmark, never as evidence for a selection rule.