00 / The research

How we got here.

Most ideas die. We publish them anyway, including the ones that did.

The research Every test, including the kills

The current ALPHAC flagship combines 3 live sleeves: AlphaMax, AlphaTrend, AlphaVintage. Their economic inputs differ, but diversification is not treated as proof of alpha. No sleeve clears the multiple-testing gate in-sample. AlphaVintage remains visible as a paper-trading history but its calendar-correct research verdict is KILLED; it is not an admitted sleeve. The live record, including a genuine risk-off episode, is the evidence that matters next.

Execution basis: paper trading and local simulation. “Live” below describes running systems, not funded execution. Model expectations are not measured forward returns. Inspect the forward evidence.

10 factors tested 2 survived 8 killed across 35+ families forward 0.3 to 0.9 (the +10% beta does NOT improve this - it adds bull upside + crash risk, not risk-adjusted quality; real chance of ~0 yr1) Sharpe grade C+

Scroll

01 / Factor tests

Every test, its real Sharpe, and its real Rank-IC.

Most ideas die. We publish every test with its real net-of-cost Sharpe and its real Rank-IC. 2 kept, 8 killed. The Rank-IC is the cross-sectional information coefficient with a Newey-West t-stat; the net Sharpe is the purged walk-forward result after the conservative cost model.

// Rank-IC is the cross-sectional information coefficient with a Newey-West t-stat. Net Sharpe is the purged walk-forward result after the cost model. Select a row for its full performance, its Rank-IC by horizon, and the source artifact.

Filter

02 / What survived

Four sleeves. One neutral core. Real, decorrelated, and honestly modest.

Market-neutral core (equity momentum + crypto funding carry + managed-futures trend + PIT macro surprise, three near-uncorrelated sleeves at equal thirds, measured average pairwise correlation +0.0345) PLUS a disclosed +10% strategic net-long overlay

AlphaMax33.3%

12-1 cross-sectional momentum, dollar-neutral long/short, split-adjusted, survivorship-free.

Standalone Sharpe+0.91

AlphaTrend33.3%

Time-series momentum across a 17-market basket (equity-index, rates, commodities, FX via liquid ETFs), long/short on each market's own trend, inverse-vol weighted. THE RE-DERIVATION IS DONE (2026-08-07) AND IT WENT AGAINST US: at the honest trial count (N=133) and the pooled trial variance, its DSR is 0.000 - not the 0.83 we published, and the WORST of the 3 sleeves rather than the best. The old 0.83 rested on a V[SR] input ~80x too small and n_trials=5 while AlphaMax was graded at N=101; it was marked on an easier exam and then called the soundest. The selected 5,133-day walk-forward artifact used for this card has net Sharpe 0.3307. A separate family-wide re-derivation has net Sharpe 0.25 over 5,179 days. Both are historical simulations, and neither clears selection adjustment. Positive skew, two-decade stable, near-uncorrelated to equities. We keep it because removing it measurably worsens book drawdown, not because its edge is established.

Standalone Sharpe+0.33

AlphaVintage33.3%

Point-in-time CPI surprise - the standardized AR(3) residual of headline (PCPI) and core (PCPIX) inflation, differenced WITHIN a single ALFRED vintage - traded as a dollar-neutral IWM-minus-SPY size spread, monthly.

Standalone Sharpe+0.23

0.035 Full-sample correlation
-- On equity down-days
-- Stress union

Measured +0.0345 average pairwise correlation across the three live sleeves (2023-07-07..2026-06-01) - POSITIVE. With the crypto sleeve, the four-sleeve v3 book measured +0.0274; suspending it raised the average. In risk-off the pairwise correlations can spike, so the diversification benefit may shrink exactly in the left tail. We size on the stressed matrix.

In-sample, research window

in-sample / research window - NOT the forward expectation

2023-07 to 2026-06 (2.9yr research window)

Sharpe
1.17
CAGR
+6.0%
Max DD
-4.6%

Honest forward expectation

After multiple-testing deflation
Sharpe
0.3 to 0.9 (the +10% beta does NOT improve this - it adds bull upside + crash risk, not risk-adjusted quality; real chance of ~0 yr1)
Return (vol-targeted)
0 to 14 (vol-targeted; lower bound is real)
Worst DD
-22 to -28 (incl. the +10% strategic-long overlay's crash tail)

The research window is short, multiple strategies were searched, and every standalone sleeve remains below the stated deflation gate. The published forward band is therefore deliberately below the in-sample headline.

Combined-book deflation

Deflation verdict
FAIL
DSR gate
0.95
Standalone equity DSR
--
Standalone crypto DSR
--

real but modest; no sleeve clears the multiple-testing deflation gate in-sample, so the only credible next evidence is a live record where the full book transacts through a risk-off episode

C+ Honest grade

real but modest; no sleeve clears the multiple-testing deflation gate in-sample, so the only credible next evidence is a live record where the full book transacts through a risk-off episode

  • Equity capacity 1B+ at Reg-T 2x gross
  • Crypto carry capacity 100k to 1M proven; decays above ~100M AUM
  • BTC-correlation cap BTC correlation caps the crypto sleeve's cross-sectional breadth, so it stays a satellite, never a scalable core.
  • Equity survivors Of the equity factors tested, only 12-1 momentum survived; value and quality did not replicate on the narrow universe.

03 / Methodology

The data layer blocks leaks at the source, not after the fact.

The validation method in plain language: point-in-time data, purged walk-forward, deflation gates, pre-registration, and a byte-reproducible backtest.

Read the execution-model boundary Engineering quality boundary Machine-readable lint debt Inspect the local execution benchmark Read the dated-futures execution foundation Read the options surface and lifecycle foundation Read the securities-borrow foundation Read the market-status and outage replay contract Read the crowding and stressed-capacity contract Read the corporate-action and delisting lifecycle contract
01

Point-in-time, survivorship-free

Every bar is stamped to the moment it was knowable; a single PIT reader enforces it on every query, so a leak is something the data layer will not permit, not something we hope to avoid.

Survivorship-free
yes
PIT enforced
yes
02

Purged walk-forward

Anchored-expanding walk-forward with a purge + embargo between train and test, so no information leaks across the boundary. Weights on a forward day are a pure function of returns strictly before that day.

Scheme
anchored-expanding, purged + embargoed
03

Deflated Sharpe Ratio

Every config that could have been measured counts toward the deflation denominator. The Deflated Sharpe Ratio (DSR) discounts the headline Sharpe by the expected maximum Sharpe of pure noise across the trial count. A clean result must clear DSR >= 0.95.

Gate
DSR >= 0.95
04

Probability of backtest overfitting

Probability of Backtest Overfitting via combinatorially-symmetric cross-validation (CSCV): the chance the in-sample best config is below-median out-of-sample. A clean result must hold PBO < 0.20.

Gate
PBO < 0.20
05

Pre-registration

The trial budget was committed to git before any sleeve touched the data lake; the commit hash is the timestamp of record. A tiny budget measured exactly once is the only honest route to a clean deflation grade.

Measure-once protocolNo re-runs, no re-windowing, no cadence search, no sign-flips, no subset search over sleeve combinations, no MVO, no static scheme in the deployable path. KILL = exclude, never re-tune.

Trial budget
9 hard ceiling
Committed
docs/design/PRE_REGISTRATION.md
06

Reproducibility

The backtest is content-hashed and byte-reproducible. A golden-master test asserts every fill price, fee and funding payment from hand-written arithmetic to 1e-9 precision. The backtest itself is not timestamp-anchored; selected heads of the separate live track-record chain have OpenTimestamps proofs.

We claim
content-hashed + byte-reproducible backtest
We do not claim
No blockchain timestamp claim for the backtest itself
Golden master
tests/integration/test_golden_master.py

04 / Crypto alone

Crypto alone is a null net of cost. Here is the proof.

Before the equity ballast, we ran crypto perpetuals through the deflation gates on their own. The result is NO-DEPLOY. The edge is too thin to survive cost.

crypto-perp strategies alone 2021-01-01 to 2026-06-01 8 trials best config A_blend

Deflated Sharpe (shared) BELOW GATE
0.2112 gate 0.95
Prob. of backtest overfitting OVER GATE
0.8818 gate < 0.20

Capacity decay, annualised Sharpe by deployed capital

  • $100kSR 0.401
  • $1MSR 0.042
  • $10MSR -0.372

The edge decays as capital grows: 0.40 Sharpe at $100k, 0.04 at $1M, and -0.37 at $10M, where market impact consumes the thin signal. Crypto carry is finite-capacity alpha, not a scalable core.

NO-DEPLOY (honest null)

Best config DSR 0.2112 < 0.95 gate AND PBO 0.8818 > 0.20 gate. No variant beats baseline and clears deflation at once. Crypto-perps alone is a thin cross-sectional signal that does not survive honest deflation net of cost; the equity sleeve is required as ballast.

05 / Research frontier

A target is not a result.

The objective is an honest FORWARD portfolio Sharpe of 2.0 across 14 evidence-backed sleeves, including at least 11 new admissions and approximately 10% maximum drawdown. This book's own measured gap between backtest and honest forward is 1.5x to 3x, so that target implies an in-sample 3.00 to 4.00. Both are published because they are different numbers. Targets never lower an admission gate.

// Feasibility comes before returns. A data-gated result is published as a constraint, not quietly removed from the queue. Credentials are requested only after a specific candidate clears its key-free evidence screen.

Read the sealed AlphaVintage missing-release correction
Book deflated Sharpe
--
PBO
< 0.20
Average correlation
--
Target breadth
14 sleeves / >= 11 new

ATLAS / Stage-zero breadth

240 coverage cells. Zero sleeves claimed.

Passing this audit is not return evidence and does not create a sleeve. It only permits a separately sealed key-free feasibility review.

Economic families
40
Coverage cells
240
Novel atlas families
16
Active feasibility
10
Retired / killed
5
Overlap review
7
Lineage audit
--
Gate evaluations
2,880
New sleeves admitted
0
Checks per survivor
86
Execution dimensions
26

60 cells may proceed only to key-free feasibility; 96 remain stopped for cited literature review; 42 require identity-level overlap resolution; 24 are retired/killed and 6 are monitoring a sealed forward-only continuation. Return data opened: 0. Return identities spent: 0.

Inspect all 240 identities Inspect all 2,880 gate evaluations Inspect source-bound lineage Inspect the admission contract

LIBRARY / Published in full

114 research documents. Published whole.

Every literature review, feasibility protocol, pre-registration and correction this programme has produced, including the ones that killed the candidate. Each is published whole, with its unrendered source beside it.

Locked return verdictKILL

earnings narrative change

Net Sharpe
-0.007
2x cost Sharpe
-0.462
DSR
0.0000
NW t-stat
-0.02
Max drawdown
-28.5%
Research gates
12/20

Not an admission. Passing this contract makes a result technically eligible for an explicit admission decision. It does not guarantee a sleeve, future performance, the target Sharpe, or the target drawdown. Point estimates cannot override confidence-bound, crisis-conditional, tail-loss, execution-evidence, or capacity-reconciliation failures. v6 additionally refuses to load if its own significance floors cannot all be satisfied at once, gates the average pairwise correlation on its upper confidence bound as well as its point estimate, and publishes the arithmetic relating the correlation gate to the portfolio objective -- including the distance the gate does not close.

01
RESEARCH QUEUED

earnings narrative change

US equities / one to three months

Point-in-time change in issuer language and disclosed operating risks, residualized against earnings surprise, momentum and sector

83,070 official 10-K records mapped; 82,491 pair-eligible filings; 0 metadata failures. The locked runner cannot load returns until the corpus is complete.

Required evidence
Point-in-time filing text, earnings-call transcripts, speaker roles, revisions and delistings
Potential sources
SEC EDGAR / LSEG StreetEvents / FactSet CallStreet
Kill rule
Kill if section lineage cannot be reconstructed point-in-time, language changes are unstable across filing formats, or residual correlation to AlphaMax exceeds 0.35.
02
DATA GATED

analyst revision drift

US equities / one to three months

Point-in-time consensus estimate revisions after controlling for earnings surprise and momentum

Required evidence
As-of analyst estimates, contributor counts, announcement timestamps and delistings
Potential sources
LSEG I/B/E/S / FactSet Estimates / Visible Alpha
Kill rule
Kill if revision timestamps cannot be reconstructed point-in-time or residual correlation to AlphaMax exceeds 0.35.
03
DATA GATED

options dispersion

equity options / weekly to monthly

Index-versus-single-stock implied correlation dispersion with defined crash loss

Required evidence
Historical SPX and constituent option chains, index membership, bid/ask, greeks and corporate actions
Potential sources
OptionMetrics / Cboe DataShop / ORATS
Kill rule
Kill if the edge is ordinary short index variance, crisis expected shortfall breaches budget, or constituent costs consume half the gross premium.

Read cited literature review

Read locked feasibility protocol

04
RESEARCH QUEUED

merger arbitrage

US equities / event driven

Announced cash-deal spread convergence conditioned on financing, regulatory and termination risk

1,965 official target anchors found. Later outcome coverage reached 91.8%, but prior deal-state coverage reached only 67.0%; the pre-written 80% gate failed. No return test was run.

A separately locked tender-only branch downloaded 100 target filings and extracted Item 4 from 94.0%. It still failed: only 10.6% yielded one safe cash price, while 74.5% were ambiguous and board posture resolved in only 22.3%. The parser was not retuned and no return test was run.

Required evidence
Point-in-time deal terms, revisions, breaks, votes, close dates and borrow costs
Potential sources
SEC EDGAR / LSEG Deals / FactSet Mergerstat
Kill rule
Kill if deal terms cannot be reconstructed without hindsight or realized break losses dominate the costed spread.
05
DATA GATED

index reconstitution flow

global equities / days to weeks

Announcement-to-effective-date demand from transparent index additions and deletions

Required evidence
Point-in-time announcement archive, index weights, float, ADV and closing-auction data
Potential sources
S&P Global / FTSE Russell / MSCI
Kill rule
Kill if announcement history is survivorship-biased or auction slippage consumes the forecast edge.
06
DATA GATED

electricity load weather spread

power and natural-gas futures / day-ahead to weekly

Forecast errors in weather-driven regional electricity load relative to marginal fuel and power prices, neutralized to commodity trend

Required evidence
Point-in-time weather forecasts, ISO load forecasts and revisions, nodal or hub power prices, fuel prices, listed contract rolls and executable bid/ask
Potential sources
EIA / NOAA / PJM / ERCOT / CME DataMine / ICE Data Services
Kill rule
Kill if forecast vintages cannot be reconstructed, listed liquidity is insufficient, or returns reduce to natural-gas trend or generic seasonality.

Read cited literature review

Read locked feasibility protocol / Inspect sealed feasibility result

07
DATA GATED

securities lending supply

US equities / days to weeks

Point-in-time borrow-fee and utilization changes as a constrained-supply return source

Required evidence
Historical borrow fees, utilization, recalls and availability by security
Potential sources
S&P Global Securities Finance / DataLend / Orbisa
Kill rule
Kill if locate and recall costs erase the signal or historical availability cannot be established point-in-time.
08
DATA GATED

credit equity relative value

corporate credit and equities / weeks to months

Issuer-level equity-versus-credit dislocation after neutralizing market, sector, duration and quality

Required evidence
Point-in-time bond prices, reference data, TRACE liquidity and equity corporate actions
Potential sources
FINRA TRACE / LSEG / Bloomberg
Kill rule
Kill if issuer mapping is hindsight-contaminated, liquidity makes fills implausible, or the result is merely quality or momentum.
09
DATA GATED

active ownership escalation

US equities / one to six months

New Schedule 13D control-intent ownership followed by specific governance, board, capital-allocation or strategic action

Schema-aware SEC discovery passed: 22,353 initial accessions, 800 immutable headers, 100% target/filer lineage, and 47.3% contemporaneous ticker coverage with a 43.8% Wilson lower bound.

The locked Item 4 stage then failed. Parser v2 resolved 159/160 submissions and extracted 139/160 sections (86.9% versus 90% required). No parser v3 or return test was opened.

Required evidence
Point-in-time Schedule 13D/13D-A state, reporting groups, Item 4 intent changes, ownership percentages, exits, delistings and borrow
Potential sources
SEC EDGAR / LSEG Ownership / FactSet Ownership
Kill rule
Kill if specific active intent cannot be separated from boilerplate point-in-time, target/reporting-person lineage is unstable, or residual correlation to existing equity sleeves exceeds 0.35.
10
RESEARCH QUEUED

treasury auction concession

US Treasury futures and liquid duration ETFs / one to three sessions

Temporary duration concession before repeated US Treasury coupon auctions followed by post-auction normalization, conditioned on issue size and dealer-balance-sheet pressure

2,037 official coupon-auction events passed the locked metadata gates, including 1,377 from 2013 onward. Core fields and event identities were complete; 99.56% had at least one day of notice. No prices were opened and zero return identities were spent.

A second timing audit found 319 two-year auctions, a median 5 calendar days of formal notice, and 0 events with the 10 calendar days needed even before counting trading sessions. The published 10-session identity remains closed until archived tentative schedules are sealed.

Required evidence
Auction announcements and results, reopening status, issue size, when-issued or futures prices, contract rolls, bid/ask and primary-dealer positioning
Potential sources
US Treasury Fiscal Data / Federal Reserve Bank of New York / CME DataMine
Kill rule
Kill if the effect is absent after the 2013 publication date, reduces to unhedged duration beta, cannot survive one-tick-plus-fee costs, or correlates above 0.35 with AlphaTrend.

Read cited literature review

Read locked feasibility protocol / Inspect sealed feasibility result

Read identity-timing protocol / Inspect identity-timing result

11
DATA GATED

cftc hedging pressure

commodity futures / weekly to monthly

Cross-sectional commodity risk transfer from commercial hedging pressure using only CFTC reports available after their Friday release, neutralized to trend and curve carry

183,676 official metadata rows across 21 years passed schema and identity checks, but exact historical release lineage is 0% and no tradable-contract map is sealed. The candidate is data-gated. No positions, prices, or returns were opened.

Required evidence
Point-in-time COT disaggregated positions, release timestamps, futures settlements, rolls, limits, bid/ask and contract multipliers
Potential sources
CFTC / CME DataMine / ICE Data Services
Kill rule
Kill if release-lagged positions do not add beyond trend and curve carry, the tradable map is unstable, crisis liquidity breaks the cost model, or correlation to AlphaTrend exceeds 0.35.

Read cited literature review

Read locked feasibility protocol / Inspect sealed feasibility result

12
DATA GATED

pre fomc announcement drift

US equity-index futures or SPY / intraday event window

A publication-aware replication of the equity premium reported in the hours before scheduled FOMC announcements, isolated from overnight and ordinary market beta

79 official regular FOMC statement events span 2016 - 2025. Release-time coverage and the 2:00 p.m. ET convention both measured 100%, but point-in-time schedule-revision lineage is 0%. Returns remain unopened and zero identities were spent.

Required evidence
Point-in-time FOMC schedule changes, intraday index futures or SPY quotes, spreads, holidays and announcement timestamps
Potential sources
Federal Reserve / Polygon / CME DataMine
Kill rule
Kill if the post-publication sample is non-positive net of costs, the result is ordinary overnight beta, schedule revisions cannot be reconstructed, or left-tail behavior violates the event-risk budget.

Read cited literature review

Read locked feasibility protocol / Inspect sealed feasibility result

13
RESEARCH QUEUED

repurchase issuance flow

US equities / one quarter to one year

Filing-known completed corporate equity demand: completed repurchases versus contamination-resolved completed issuance, never authorization headlines

SEC COLLECTION AND SEMANTICS COMPLETE AWAITING BLIND LABELS. The 48-test no-return pipeline is implemented. Every stage binds exact manifest identities, immutable part hashes and zero return access, and exits nonzero when its required seal or gate is absent. Blind labels must be sealed against the parser source before parser output exists. Company Facts or Item 703 extraction alone can never pass full feasibility; combined coverage, amendment replay, quarterization and contamination gates remain mandatory.

Required evidence
PIT XBRL facts, accession-bound Item 703 tables, shares and corporate actions, issuance contamination flags, delistings, borrow, spreads and market impact
Potential sources
SEC EDGAR / LSEG Fundamentals / FactSet Fundamentals
Kill rule
Kill if completed repurchases and issuance cannot be separated point-in-time from authorizations, amendments, stock compensation, splits, acquisitions and custom tags, or if a later residual is subsumed by existing equity sleeves.

Read cited literature review

Read locked feasibility protocol

06 / Roadmap

The binding constraint is sample length, not idea count.

The binding constraint is trustworthy evidence, not idea count. The current four-sleeve core must earn its record forward while new candidates move through one-shot, pre-registered gates. The honest path is three-step.

  1. 01

    Earn the current book forward

    Accrue genuinely out-of-sample paper evidence on the four equal-quarter neutral sleeves, with the +10% strategic beta overlay reported separately. Do not relabel simulations as realized performance or rescue a weak result with an unregistered configuration change.

  2. 02

    Execute the registered discovery queue

    Test differentiated candidates across narrative, event, options, lending, power, credit, and ownership data under point-in-time, cost, capacity, deflation, and correlation gates. Publish ADD, KILL, or DATA-ESCALATE; never tune a failed one-shot identity after seeing returns.

  3. 03

    Buy data and infrastructure only after evidence

    Escalate to historical borrow/locate, options, broker, execution, or other paid credentials only when a named candidate has cleared every key-free screen. Credentials such as Alpaca belong to approved shadow execution, not to exploratory backtest rescue.

// Phase-8 pre-arm gates (C3 / C5 / C7 / C10) must clear before live capital.

This export

Schema
glassbox.research/1
Generated
2026-11-10 01:27:55 UTC
Content hash
sha256:cd0708826cc0...b4ee50

// Every published number is read from a real engine artifact. Where a value does not exist in an artifact it is omitted, never invented. The forward Sharpe is the deflated 0.3-0.9 expectation, never the in-sample headline; killed factors carry their real negative net Sharpes.

Methodology and questions Read every measurement in full Inspect every registered trial See it proven in the open Join the waitlist