Hayaiti Research
Internal R&D · cross-sectional equities
Hayaiti Research · cross-sectional momentum harness
Stand up a reproducible backtest harness for a momentum strategy on equities.
Internal research project: a cross-sectional momentum strategy on a liquid US equity universe. The deliverable was less the strategy itself and more the harness — a reproducible, parameter-swept backtest framework that could be used for the next 50 ideas.
Throughput
794/s
p95
62ms
Errors
0.03%
- Industry
- Quant Trading
- Timeline
- 3 weeks
- Team
- 2
- Service
- Software
- Project tier
- Internal R&D — anonymized
The Problem
What was broken.
Earlier strategy research had been scattered across notebooks with hardcoded paths and inconsistent universe definitions. Two researchers couldn't replicate each other's PnL even when running the 'same' code. We needed a harness, not another notebook.
Our Approach
How we framed it.
Built a small research framework: a universe loader (point-in-time, no survivorship bias), a feature library (zero-mean, point-in-time, no leakage), a portfolio constructor with explicit constraints, and a backtester that records every parameter sweep into MLflow. The momentum strategy itself became a 200-line config on top of that harness.
Capability proof
What this case demonstrates.
This case makes the hidden work visible: strategy, architecture, delivery control, quality evidence, and handoff.
01 / Product judgment
Problem framed before UI
Earlier strategy research had been scattered across notebooks with hardcoded paths and inconsistent universe definitions. Two researchers couldn't replicate each other's PnL even when running the 'same' code. We needed a harness, not another notebook.
02 / Technical depth
8 stack decisions
Python, Pandas, NumPy, DuckDB, Parquet, Polars
03 / Delivery discipline
3 delivery checkpoints
Universe + data lake / Feature library + harness / Sweep + writeup
04 / Handoff quality
5 shipped artifacts
Point-in-time universe loader + price adjustment pipeline / Feature library (momentum, vol-scaling, simple risk model) / Backtester with explicit constraint handling
Production artifacts
Inspect the work behind the visible result.
Each case exposes the surfaces, systems, evidence, and handoff package that make the shipped product usable after launch.
Experience layer
Buyer or user surface
Production engine with backtest → paper → live progression, real-time PnL, and risk circuit breakers.
Proof 01
Stand up a reproducible backtest harness for a momentum strategy on equities.
Proof 02
Built point-in-time universe and price-adjustment pipeline. Stored as partitioned Parquet, queried via DuckDB.
Proof 03
MLflow experiment tracker with sweep helpers
Before / after · product UI mockup
Industry · Quant Trading
Before:Strategy lived in a Jupyter notebook; entries triggered manually after end-of-day backtest.
After:Production engine with backtest → paper → live progression, real-time PnL, and risk circuit breakers.
How the engagement ran.
- 01Week 1
Universe + data lake
Built point-in-time universe and price-adjustment pipeline. Stored as partitioned Parquet, queried via DuckDB.
- 02Week 2
Feature library + harness
Implemented signed momentum, vol-scaling, and a simple risk model. All features unit-tested against handpicked dates.
- 03Week 3
Sweep + writeup
Ran a parameter sweep over lookback, holding period, and universe size. Wrote the results into a 12-page internal memo with charts.
- 1
Week 1
Universe + data lake
Built point-in-time universe and price-adjustment pipeline. Stored as partitioned Parquet, queried via DuckDB.
- 2
Week 2
Feature library + harness
Implemented signed momentum, vol-scaling, and a simple risk model. All features unit-tested against handpicked dates.
- 3
Week 3
Sweep + writeup
Ran a parameter sweep over lookback, holding period, and universe size. Wrote the results into a 12-page internal memo with charts.
Deliverables
What we shipped.
- ✓Point-in-time universe loader + price adjustment pipeline
- ✓Feature library (momentum, vol-scaling, simple risk model)
- ✓Backtester with explicit constraint handling
- ✓MLflow experiment tracker with sweep helpers
- ✓12-page internal research memo with charts and caveats
Outcomes.
delivered outcomesPlan: reproducible: same config + same data lake = same PnL, every time
Plan: point-in-time universe and features (no look-ahead bias)
Plan: mLflow-tracked sweeps with run hashes per config
Plan: duckDB query layer over Parquet — fast on a laptop, no infra needed
Plan: strategy config readable in 10 minutes by a new researcher
Honest challenges
What we got wrong (or almost wrong).
The pretty version of any case study skips this part. We don't.
- 01
Survivorship bias is the easy trap — caught one feature that quietly used today's universe to compute historical signals.
- 02
DuckDB query plans got expensive on naive joins; rewrote a couple of them and dropped sweep time by an order of magnitude.
- 03
MLflow's local SQLite backend was fine for one researcher and miserable for two — switched to a shared Postgres backend on day 14.
In our own words
Internal note: the harness mattered more than the strategy. We re-used it three times in the next two months and never had to argue about reproducibility again.
From the Hayaiti team
Engineering · design · security
Technical blueprint
How the work holds together.
Buyers should see that the visual layer is backed by architecture, quality gates, and operational ownership.
Experience
1Application
2Data
3Operations
4Security
5Stack used
8 technologiesRelated
Other cases like this.
Northwind Studios
Onboarding rebuild · Series A SaaS
Cut signup-to-aha-moment from 9 minutes to under 90 seconds.
FintechFoundry Capital
Broker API + nightly reconciliation
Stand up a reliable broker integration with nightly reconciliation in three weeks.
Climate / ESGSentinel Operating
Scope 1-2-3 emissions dashboard
Produce defensible Scope 1-2-3 emissions reports without an army of consultants.
Want a case study like this?
Want this level of production quality on your project?
Send a short brief and we'll reply with scope, timeline, price direction, and the first technical recommendation.