X_Quant: A Quant Research Program That Learned to Say No

Correction (October 2026). Earlier versions of this page presented a frozen momentum strategy (cfg_00 / MetaEnsemble v56) as CSCV-validated and trading on a paper account, with headline overfitting and out-of-sample metrics. That system was rejected on 2 August 2026 and decommissioned on 9 August 2026. The metrics previously shown here are withdrawn and should not be cited.

Overview

X_Quant is my solo quantitative-research program. It has had two lives.

v1 grew over two years into a ≈274K-line trading platform built around a single strategy. It was rejected on 2 August 2026 and decommissioned on 9 August. v2 started a week later — the day v1 was decommissioned — from zero, with a written research constitution drawn from v1's post-mortem. It then spent seven weeks measuring, under that constitution, whether any strategy tradeable at retail scale had a real edge. None did.

v1 (2024–2026): lessons the post-mortem left behind

The v2 documents do not record a single verdict on v1. What they record are the lessons they attribute to it:

• PBO had been computed in-sample. v2's standard: "PBO only on the out-of-sample matrix — PBO on in-sample data is not PBO."
• Free-form feature discovery. v2's standard calls it "an overfitting machine" (a v1 lesson) and allows machine learning only to combine at most three pre-registered signals.
• An unverified cost assumption. v1's paper-fill cost estimate was never verified.
• Eleven engineering failure patterns, each now paired with a guard rule: code and tests reading dead copies of the tree; a monitor that overloaded the very resource it was watching; gates that reported "clean" when they had not measured anything; empty results read as success; tests that wrote to the live journal; multi-hour runs with no checkpoints; evidence kept outside version control — and four more.

The through-line: v1 had validation machinery, but not a validation process that could stop me from believing my own numbers.

v2 (August–September 2026): a process built to kill bad ideas cheaply

v2 is governed by a written constitution of eight articles. A result produced in breach of any of them is void, however good it looks. The core of it:

• Truth of measurement — causality checked per bar; PBO only out-of-sample; folds that cover 2008, 2011, 2020 and 2022; pre-registration before the run, one shot; data pinned by hashed manifests.
• The instrument is not the strategy — the measurement harness lives apart from strategies; repair and verify run in separate sessions, and no session judges its own repair.
• Limits are law — every run has a wall-clock budget; gates fail closed; zero means "not measured", never "success".
• A register of truth — an append-only decision register; a wrong entry is corrected by a new entry, never by editing the old one.
• Edge before code — no strategy without a written, falsifiable hypothesis: who pays, why it isn't arbitraged away, and how its death will be recognised.
• Honest data from day zero — total-return prices and point-in-time constituents; survivorship-biased data is not data.
• One engine for research and production, with bit-for-bit parity.

Around that sit a permanent attempts ledger (every configuration ever evaluated counts toward the deflated Sharpe of every future test — 2,976 attempts at the last committed run), a sealed 2018–2025 holdout for the equity discovery programmes that only a pre-registered one-shot confirmation could open (none ever qualified), an incident log where an incident closes only when its root cause is removed, and decisions taken in a recorded dialogue with an adversarial AI partner (Claude) whose objections stay in the register.

Results on record

Line of work	What was tested	Recorded verdict
Post-earnings drift, small/micro caps	Pre-registered event hypothesis	Dead — gross drift +0.26%, net −0.74% against a pre-registered +1.5% bar
Insider-purchase clusters	Pre-registered event hypothesis	Dead
8-K filing drift	Pre-registered event hypothesis	Dead
Event-signal discovery scan	633 configurations	0 passed all three stages
Signal-combination portfolio	57 weak signals combined	Not a candidate — gross Sharpe +2.33, net −2.46
Indicator discovery scan	166 configurations	0 passed the first stage
Re-costing at 5k / 25k / 100k USD	The 57 first-stage signals	0 survived costs at any capital level
Replication of Carver's trend system on ETFs	Three versions (below)	Closed — no measurable edge at retail scale
NQ futures track	Third-party rules, ensembles, ML (below)	No tradeable candidate

By 23 September, in the register's words (translated from Polish): "After 7 weeks and 8 programmes: 0 candidates." The futures track that followed added none. Nothing in v2 was ever recorded as trading on paper or live.

The Carver replication

The deepest line of work replicated a published trend-following system on ETFs, each version under its own specification:

• v1.0 (85 ETFs, six rules) reached a gross Sharpe of +1.00, in line with the author's published figures — but was −1.46 net at 25k: small orders cannot carry retail costs.
• v1.1, a dry-run of the author's own cost rule (its spec deliberately never frozen): 0 of 528 rule–instrument pairs were admissible at any capital up to 50k.
• v1.2.1 was fitted to a small retail account (four ETFs, two rules), calibrated against a martingale first, then run once: gross −0.07, net −0.14, at the 45th percentile of the null distribution — indistinguishable from no signal.

A separate verify session reproduced the verdict from the committed artifacts. It found a wrong number in the register itself (positive years: 4, not 3) plus several smaller errata; all were corrected by a new entry, and the standard was tightened so that a number may enter the register only from a numeric artifact, never from report prose. The conclusion that closed the line: the author's edge is breadth, and breadth under his cost rule needs at least ~100k on ETFs.

The NQ futures track

In the last week, a futures track worked through a public third-party rule set and the strategies built around it. These are build-session results; none was independently verified:

• A public third-party NQ rule set was reproduced: the replay matched its published ledger, which confirmed the numbers but did not demonstrate an edge. It received no pass and served, unchanged, as the baseline for the tests below.
• A one-shot out-of-sample test of an ensemble of its variants: Sharpe 0.19, profit factor 1.016 — failed four of five criteria.
• An ensemble with an in-sample Sharpe of 2.84 fell to −0.88 and −0.48 on clean out-of-sample windows — rejected as overfit.
• A second ensemble passed its out-of-sample test formally (Sharpe 1.18), but its gain came from the window in which its families were discovered — its pass was treated as void, and the paper forward test proposed for it never produced a recorded result.
• A walk-forward of the whole selection procedure over 2014–2026 reached Sharpe 0.36 against 0.8 and PSR 90.2% against 95% — failed, with no threshold loosened.
• A pooled gradient-boosting model across seven markets — a departure from the "combine at most three pre-registered signals" standard — reached a walk-forward Sharpe of −0.42 and failed, with no tuning and no second attempt allowed.

One last micro-futures run (29 September) was executed without a freeze step and its results were never committed, so under v2's own rules it has no result. A prediction-market track got as far as a design package; no model was built.

What it produced

The 23 September assessment that closed the ETF and equity phase names the real output: not a strategy, but the instrument — an engine, a register, a calibration method and a set of executable gates that can close a research question with evidence in about four sessions. That instrument, and the habit of not believing my own numbers until something independent has tried to break them, is what carried over.

Status

No commits or register entries since 29 September 2026; neither version's repository is public. Its successor is Alpha Lab, which reads X_Quant's data read-only and deliberately starts its own attempts ledger from zero.

Tech Stack

Python, pandas, NumPy, PyArrow / Parquet, LightGBM, pytest with mutation tests on the gates, hashed data manifests, git-pinned pre-registrations, and append-only Markdown registers.