Writing

Notes from the loop.

Short pieces on how I think about research-engineering for systematic trading. Practical, opinionated, sized for an evening's commute.

Start here

Read these three in order for a quick orientation to how this site thinks about research-engineering for systematic trading.

01 · intro · 3 min

MLOps for quant research isn't MLOps for ML

Why the MLOps vocabulary you know doesn't quite fit when the model trades.
02 · intro · 6 min

Right-sizing the research data platform: five thresholds, in order

How to pick the smallest research-data architecture that fits your team's current size — and when to upgrade.
03 · intro · 4 min

The deploy contract isn't a YAML file

Why the 3–4 promises a quant makes before going to production matter more than how the framework enforces them.

tag filter, search, and keyboard shortcuts need JS. all posts are listed below — RSS at /feed.xml.

Citation graph

Forward edges from `cites:` frontmatter (validated at build time). Reverse edges are the same data, indexed by target — read across the row to follow a thesis backward.

Note	Builds on	Cited by (writing)	Cited by (outside)
AI research agents as platform citizens	Walk-forward without leakage: a checklist that's saved me The promotion gate: why bad data should be unreachable The deploy contract isn't a YAML file	Refusal as planning hint The citation graph as substrate	agent · 019e4900 agent · 019e4900
An event-aware wrapper needs signal concentration, not just signal presence	The baseline arm you forgot to include The event clock isn't the panel clock Don't pay for caution you can't justify	Designing an event-aware strategy: a checklist	—
Anti-correlation hedges, doesn't stack — the pairwise rule's 5th iteration	The pairwise rule — final form (four iterations later) The full pairwise matrix — 28 arms at N=100 Dual-signal at N=100: the composite stacks (slightly)	The common-clock replay: make two strategies comparable before you believe their sum What 29 arms told us — session research log v4	—
Audit isn't a feature you turn on	The citation graph as substrate Refusal as planning hint The promotion gate: why bad data should be unreachable	The common-clock replay: make two strategies comparable before you believe their sum	—
Composite strategies can interfere	Three-clock momentum tops the harness Vol ratio as cross-symbol regime gate Vol-regime filter wins on mean, pays in variance	Dual-signal data makes composites stack Pairwise correlation predicts composition before you run it Perfect correlation explains the interference The strategy catalog — twenty-nine panel/event shapes, one ranking helper What 19 arms told us about strategy composition	—
Designing an event-aware strategy: a checklist	The event clock isn't the panel clock Don't pay for caution you can't justify The baseline arm you forgot to include An event-aware wrapper needs signal concentration, not just signal presence	Drift vs reversal: the cleanest counterfactual for a post-event regime Event gates cost Sharpe when the event has edge Higher moments add noise faster than signal Property tests catch cross-strategy bugs that per-strategy tests miss Strategy shape beats factor count: TSMom > 4-factor on the same data The strategy catalog — twenty-nine panel/event shapes, one ranking helper	—
Disconfirmed: the transition-gate fix didn't recover the regime-filter	Vol-clustering breaks the regime filter — but not for the obvious reason Vol ratio as cross-symbol regime gate Vol-regime filter wins on mean, pays in variance	Three vol-intervention experiments, zero wins on this synthetic	—
Don't pay for caution you can't justify	The event clock isn't the panel clock Walk-forward without leakage: a checklist that's saved me	Designing an event-aware strategy: a checklist Drift vs reversal: the cleanest counterfactual for a post-event regime An event-aware wrapper needs signal concentration, not just signal presence The baseline arm you forgot to include The strategy catalog — twenty-nine panel/event shapes, one ranking helper	—
Drift vs reversal: the cleanest counterfactual for a post-event regime	The event clock isn't the panel clock The baseline arm you forgot to include Don't pay for caution you can't justify Designing an event-aware strategy: a checklist	Event gates cost Sharpe when the event has edge	—
Dual-signal at N=100, all 27 arms — the full picture	One hundred seeds confirms and converges Dual-signal at N=100: the composite stacks (slightly) What 27 arms told us — session research log v3	The full pairwise matrix — 28 arms at N=100	—
Dual-signal at N=100: the composite stacks (slightly)	The pairwise rule — final form (four iterations later) One hundred seeds confirms and converges Dual-signal data makes composites stack	Anti-correlation hedges, doesn't stack — the pairwise rule's 5th iteration Dual-signal at N=100, all 27 arms — the full picture What 27 arms told us — session research log v3	—
Dual-signal data makes composites stack	Composite strategies can interfere Vol-regime filter wins on mean, pays in variance Three-clock momentum tops the harness	Dual-signal at N=100: the composite stacks (slightly) Pairwise correlation predicts composition before you run it The strategy catalog — twenty-nine panel/event shapes, one ranking helper What 19 arms told us about strategy composition	—
Event gates cost Sharpe when the event has edge	The baseline arm you forgot to include Designing an event-aware strategy: a checklist Drift vs reversal: the cleanest counterfactual for a post-event regime Synthetic data shows what you were solving for Long-only buys asymmetric exposure, not just lower Sharpe	The single-seed lead was a fluke	—
Fifty seeds reveal the tie	The pairwise rule predicted this one What 24 arms told us — the session's research log Vol-regime filter wins on mean, pays in variance	Fifty seeds, twenty-four arms — the full leaderboard One hundred seeds confirms and converges The clustered-vol finding was also small-N — another tie at higher N	—
Fifty seeds, twenty-four arms — the full leaderboard	Fifty seeds reveal the tie What 24 arms told us — the session's research log Strategy shape beats factor count: TSMom > 4-factor on the same data	One hundred seeds confirms and converges	—
Higher moments add noise faster than signal	The baseline arm you forgot to include The strategy catalog — twenty-nine panel/event shapes, one ranking helper Designing an event-aware strategy: a checklist	Strategy shape beats factor count: TSMom > 4-factor on the same data Vol-of-vol distinguishes regime from level	—
How to use the comparison harness	The strategy catalog — twenty-nine panel/event shapes, one ranking helper What 19 arms told us about strategy composition Pairwise correlation predicts composition before you run it	—	—
Long-only buys asymmetric exposure, not just lower Sharpe	The strategy catalog — twenty-nine panel/event shapes, one ranking helper Strategy shape beats factor count: TSMom > 4-factor on the same data	Event gates cost Sharpe when the event has edge Sharpe is scale-invariant. Stop trying to make it not. The single-seed lead was a fluke Three-clock momentum tops the harness Vol-regime filter wins on mean, pays in variance	—
MLOps for quant research isn't MLOps for ML	—	—	agent · 019e4900
Observability for alpha pipelines: three dashboards, one rule	—	—	agent · 019e4900 agent · 019e4900
One hundred seeds confirms and converges	Fifty seeds, twenty-four arms — the full leaderboard Fifty seeds reveal the tie What 24 arms told us — the session's research log	Dual-signal at N=100, all 27 arms — the full picture Dual-signal at N=100: the composite stacks (slightly) The clustered-vol finding was also small-N — another tie at higher N The first robust single-signal stack What 27 arms told us — session research log v3 Why ts_momentum leads at N=100	—
Pairwise correlation predicts composition before you run it	Composite strategies can interfere Dual-signal data makes composites stack What 19 arms told us about strategy composition	How to use the comparison harness Perfect correlation explains the interference The pairwise rule — final form (four iterations later) The pairwise rule predicted this one	—
Perfect correlation explains the interference	Composite strategies can interfere Pairwise correlation predicts composition before you run it The intervention-point rule, confirmed by a fourth experiment	The sign-vs-rank composition conflict The pairwise rule — final form (four iterations later) The pairwise rule predicted this one What 24 arms told us — the session's research log	—
Property tests catch cross-strategy bugs that per-strategy tests miss	The strategy catalog — twenty-nine panel/event shapes, one ranking helper Designing an event-aware strategy: a checklist	—	—
Redundant vs multiplicative composition — what +0.70 means	The first robust single-signal stack The pairwise rule — final form (four iterations later) The full pairwise matrix — 28 arms at N=100	—	—
Refusal as planning hint	AI research agents as platform citizens The citation graph as substrate	Audit isn't a feature you turn on Search isn't research	—
Right-sizing the research data platform: five thresholds, in order	—	—	agent · 019e4900
Search isn't research	The promotion gate: why bad data should be unreachable Refusal as planning hint	—	—
Sharpe is scale-invariant. Stop trying to make it not.	Long-only buys asymmetric exposure, not just lower Sharpe The strategy catalog — twenty-nine panel/event shapes, one ranking helper	—	—
Six regression tests pin the findings	What 29 arms told us — session research log v4 The pairwise rule — final form (four iterations later)	—	—
Strategy shape beats factor count: TSMom > 4-factor on the same data	The strategy catalog — twenty-nine panel/event shapes, one ranking helper Higher moments add noise faster than signal Designing an event-aware strategy: a checklist	Fifty seeds, twenty-four arms — the full leaderboard Long-only buys asymmetric exposure, not just lower Sharpe Synthetic data shows what you were solving for The strategy catalog — twenty-nine panel/event shapes, one ranking helper Three-clock momentum tops the harness What 19 arms told us about strategy composition Why ts_momentum leads at N=100	—
Synthetic data shows what you were solving for	The baseline arm you forgot to include Strategy shape beats factor count: TSMom > 4-factor on the same data	Event gates cost Sharpe when the event has edge The single-seed lead was a fluke	—
The baseline arm you forgot to include	Don't pay for caution you can't justify Walk-forward without leakage: a checklist that's saved me The event clock isn't the panel clock	Designing an event-aware strategy: a checklist Drift vs reversal: the cleanest counterfactual for a post-event regime An event-aware wrapper needs signal concentration, not just signal presence Event gates cost Sharpe when the event has edge Higher moments add noise faster than signal Synthetic data shows what you were solving for The single-seed lead was a fluke	—
The citation graph as substrate	The promotion gate: why bad data should be unreachable AI research agents as platform citizens	Audit isn't a feature you turn on Refusal as planning hint The event clock isn't the panel clock	agent · 019e4900 agent · 019e4900
The clustered-vol finding was also small-N — another tie at higher N	Vol-clustering breaks the regime filter — but not for the obvious reason Fifty seeds reveal the tie One hundred seeds confirms and converges	What 27 arms told us — session research log v3	—
The common-clock replay: make two strategies comparable before you believe their sum	Audit isn't a feature you turn on Anti-correlation hedges, doesn't stack — the pairwise rule's 5th iteration	—	—
The deploy contract isn't a YAML file	—	AI research agents as platform citizens The strategy catalog — twenty-nine panel/event shapes, one ranking helper	agent · 019e4900
The event clock isn't the panel clock	Walk-forward without leakage: a checklist that's saved me The citation graph as substrate	Designing an event-aware strategy: a checklist Don't pay for caution you can't justify Drift vs reversal: the cleanest counterfactual for a post-event regime An event-aware wrapper needs signal concentration, not just signal presence The baseline arm you forgot to include The strategy catalog — twenty-nine panel/event shapes, one ranking helper	—
The first robust single-signal stack	The full pairwise matrix — 28 arms at N=100 What 29 arms told us — session research log v4 One hundred seeds confirms and converges	Redundant vs multiplicative composition — what +0.70 means	—
The full pairwise matrix — 28 arms at N=100	The pairwise rule — final form (four iterations later) Dual-signal at N=100, all 27 arms — the full picture What 27 arms told us — session research log v3	Anti-correlation hedges, doesn't stack — the pairwise rule's 5th iteration Redundant vs multiplicative composition — what +0.70 means The first robust single-signal stack What 29 arms told us — session research log v4	—
The intervention-point rule, confirmed by a fourth experiment	Three vol-intervention experiments, zero wins on this synthetic Vol-clustering breaks the regime filter — but not for the obvious reason Vol ratio as cross-symbol regime gate	Perfect correlation explains the interference What 24 arms told us — the session's research log	—
The pairwise rule — final form (four iterations later)	Pairwise correlation predicts composition before you run it Perfect correlation explains the interference The pairwise rule predicted this one The sign-vs-rank composition conflict	Anti-correlation hedges, doesn't stack — the pairwise rule's 5th iteration Dual-signal at N=100: the composite stacks (slightly) Redundant vs multiplicative composition — what +0.70 means Six regression tests pin the findings The full pairwise matrix — 28 arms at N=100 What 27 arms told us — session research log v3	—
The pairwise rule predicted this one	Perfect correlation explains the interference Pairwise correlation predicts composition before you run it Three-clock momentum tops the harness	Fifty seeds reveal the tie The sign-vs-rank composition conflict The pairwise rule — final form (four iterations later) What 24 arms told us — the session's research log	—
The promotion gate: why bad data should be unreachable	—	AI research agents as platform citizens Audit isn't a feature you turn on Search isn't research The citation graph as substrate	/sandbox/portfolio agent · 019e4900 agent · 019e4900
The sign-vs-rank composition conflict	The pairwise rule predicted this one Why ts_momentum leads at N=100 Perfect correlation explains the interference	The pairwise rule — final form (four iterations later)	—
The single-seed lead was a fluke	Event gates cost Sharpe when the event has edge The baseline arm you forgot to include Long-only buys asymmetric exposure, not just lower Sharpe Synthetic data shows what you were solving for	—	—
The strategy catalog — twenty-nine panel/event shapes, one ranking helper	The deploy contract isn't a YAML file The event clock isn't the panel clock Don't pay for caution you can't justify Designing an event-aware strategy: a checklist Strategy shape beats factor count: TSMom > 4-factor on the same data Composite strategies can interfere Dual-signal data makes composites stack	Higher moments add noise faster than signal How to use the comparison harness Long-only buys asymmetric exposure, not just lower Sharpe Property tests catch cross-strategy bugs that per-strategy tests miss Sharpe is scale-invariant. Stop trying to make it not. Strategy shape beats factor count: TSMom > 4-factor on the same data Three-clock momentum tops the harness Vol-of-vol distinguishes regime from level Vol ratio as cross-symbol regime gate What 19 arms told us about strategy composition	—
Three vol-intervention experiments, zero wins on this synthetic	Vol-clustering breaks the regime filter — but not for the obvious reason Disconfirmed: the transition-gate fix didn't recover the regime-filter What 19 arms told us about strategy composition	The intervention-point rule, confirmed by a fourth experiment What 24 arms told us — the session's research log	—
Three-clock momentum tops the harness	The strategy catalog — twenty-nine panel/event shapes, one ranking helper Strategy shape beats factor count: TSMom > 4-factor on the same data Long-only buys asymmetric exposure, not just lower Sharpe	Composite strategies can interfere Dual-signal data makes composites stack The pairwise rule predicted this one Vol-regime filter wins on mean, pays in variance What 19 arms told us about strategy composition	—
Vol ratio as cross-symbol regime gate	Vol-of-vol distinguishes regime from level The strategy catalog — twenty-nine panel/event shapes, one ranking helper	Composite strategies can interfere Disconfirmed: the transition-gate fix didn't recover the regime-filter The intervention-point rule, confirmed by a fourth experiment Vol-clustering breaks the regime filter — but not for the obvious reason Vol-regime filter wins on mean, pays in variance	—
Vol-clustering breaks the regime filter — but not for the obvious reason	Vol ratio as cross-symbol regime gate Vol-regime filter wins on mean, pays in variance What 19 arms told us about strategy composition	Disconfirmed: the transition-gate fix didn't recover the regime-filter The intervention-point rule, confirmed by a fourth experiment The clustered-vol finding was also small-N — another tie at higher N Three vol-intervention experiments, zero wins on this synthetic	—
Vol-of-vol distinguishes regime from level	The strategy catalog — twenty-nine panel/event shapes, one ranking helper Higher moments add noise faster than signal	Vol ratio as cross-symbol regime gate	—
Vol-regime filter wins on mean, pays in variance	Vol ratio as cross-symbol regime gate Three-clock momentum tops the harness Long-only buys asymmetric exposure, not just lower Sharpe	Composite strategies can interfere Disconfirmed: the transition-gate fix didn't recover the regime-filter Dual-signal data makes composites stack Fifty seeds reveal the tie Vol-clustering breaks the regime filter — but not for the obvious reason What 19 arms told us about strategy composition	—
Walk-forward without leakage: a checklist that's saved me	—	AI research agents as platform citizens Don't pay for caution you can't justify The baseline arm you forgot to include The event clock isn't the panel clock	/sandbox/microstructure /sandbox/portfolio agent · 019e4900 agent · 019e4900 agent · 019e4900
What 19 arms told us about strategy composition	The strategy catalog — twenty-nine panel/event shapes, one ranking helper Strategy shape beats factor count: TSMom > 4-factor on the same data Three-clock momentum tops the harness Vol-regime filter wins on mean, pays in variance Composite strategies can interfere Dual-signal data makes composites stack	How to use the comparison harness Pairwise correlation predicts composition before you run it Three vol-intervention experiments, zero wins on this synthetic Vol-clustering breaks the regime filter — but not for the obvious reason What 24 arms told us — the session's research log	—
What 24 arms told us — the session's research log	What 19 arms told us about strategy composition Three vol-intervention experiments, zero wins on this synthetic The intervention-point rule, confirmed by a fourth experiment Perfect correlation explains the interference The pairwise rule predicted this one	Fifty seeds, twenty-four arms — the full leaderboard Fifty seeds reveal the tie One hundred seeds confirms and converges What 27 arms told us — session research log v3 Why ts_momentum leads at N=100	—
What 27 arms told us — session research log v3	What 24 arms told us — the session's research log The pairwise rule — final form (four iterations later) Dual-signal at N=100: the composite stacks (slightly) One hundred seeds confirms and converges The clustered-vol finding was also small-N — another tie at higher N	Dual-signal at N=100, all 27 arms — the full picture The full pairwise matrix — 28 arms at N=100 What 29 arms told us — session research log v4	—
What 29 arms told us — session research log v4	What 27 arms told us — session research log v3 Anti-correlation hedges, doesn't stack — the pairwise rule's 5th iteration The full pairwise matrix — 28 arms at N=100	Six regression tests pin the findings The first robust single-signal stack	—
Why ts_momentum leads at N=100	One hundred seeds confirms and converges Strategy shape beats factor count: TSMom > 4-factor on the same data What 24 arms told us — the session's research log	The sign-vs-rank composition conflict	—

Notes from the loop.

Start here

MLOps for quant research isn't MLOps for ML

Right-sizing the research data platform: five thresholds, in order

The deploy contract isn't a YAML file

More

The common-clock replay: make two strategies comparable before you believe their sum

Anti-correlation hedges, doesn't stack — the pairwise rule's 5th iteration

Composite strategies can interfere

Designing an event-aware strategy: a checklist

Disconfirmed: the transition-gate fix didn't recover the regime-filter

Don't pay for caution you can't justify

Drift vs reversal: the cleanest counterfactual for a post-event regime

Dual-signal at N=100, all 27 arms — the full picture

Dual-signal data makes composites stack

Dual-signal at N=100: the composite stacks (slightly)

An event-aware wrapper needs signal concentration, not just signal presence

Event gates cost Sharpe when the event has edge

Fifty seeds, twenty-four arms — the full leaderboard

Fifty seeds reveal the tie

Higher moments add noise faster than signal

How to use the comparison harness

One hundred seeds confirms and converges

The intervention-point rule, confirmed by a fourth experiment

Long-only buys asymmetric exposure, not just lower Sharpe

Pairwise correlation predicts composition before you run it

Perfect correlation explains the interference

Property tests catch cross-strategy bugs that per-strategy tests miss

Redundant vs multiplicative composition — what +0.70 means

Sharpe is scale-invariant. Stop trying to make it not.

The sign-vs-rank composition conflict

Six regression tests pin the findings

Strategy shape beats factor count: TSMom > 4-factor on the same data

Synthetic data shows what you were solving for

The baseline arm you forgot to include

The clustered-vol finding was also small-N — another tie at higher N

The event clock isn't the panel clock

The first robust single-signal stack

The full pairwise matrix — 28 arms at N=100

The pairwise rule — final form (four iterations later)

The pairwise rule predicted this one

The single-seed lead was a fluke

The strategy catalog — twenty-nine panel/event shapes, one ranking helper

Three-clock momentum tops the harness

Three vol-intervention experiments, zero wins on this synthetic

Vol-clustering breaks the regime filter — but not for the obvious reason

Vol-of-vol distinguishes regime from level

Vol ratio as cross-symbol regime gate

Vol-regime filter wins on mean, pays in variance

What 19 arms told us about strategy composition

What 24 arms told us — the session's research log

What 27 arms told us — session research log v3

What 29 arms told us — session research log v4

Why ts_momentum leads at N=100

AI research agents as platform citizens

Audit isn't a feature you turn on

Refusal as planning hint

Search isn't research

The citation graph as substrate

The promotion gate: why bad data should be unreachable

Walk-forward without leakage: a checklist that's saved me

Observability for alpha pipelines: three dashboards, one rule

Citation graph