Research engineering

The system behind the research.

A model is one component. The surrounding work—data quality, timing, tests and replayability—is what makes an experiment interpretable.

Python research workflowsPoint-in-time thinkingExplicit evidence boundaries

Architecture, with a purpose

Built around
the decision.

Monolith’s research work connects market data, experimental models, portfolio logic and evidence records. Each layer answers a different question.

Source → observable information
Identify the origin, timestamp and coverage of each input. An observation date is not necessarily the date a researcher could have known it.
Transform → explicit features
Build alignment and feature rules that can be inspected. Treat revisions, missing data and universe membership as modelling inputs.
Experiment → bounded question
Separate formation and evaluation. Specify benchmarks, costs and the unit of observation before interpreting a result.
Record → reproducible decision
Keep code, outputs and limitations together. A readable receipt should distinguish what executed from what remains assumed.

01Engineering problems that matter

Time is part of the data.

Public availability, vendor ingestion and the strategy’s decision clock must not be collapsed into a single date.

PIT

Missing is not zero.

An unavailable observation is not the same as a verified empty record or an observed zero return.

DATA

Tests are not alpha.

Software tests can validate an implementation boundary. They cannot establish a market effect by themselves.

TEST

Research is not execution.

A successful replay does not prove broker fills, financing, capacity or real-money performance.

OPS

A practical
research toolkit.

PythonPandas / NumPySQLGitAutomated tests

The existing research corpus includes statistical-arbitrage engines, macro-regime models, equity-factor work and alternative-data studies. The useful common ground is not a single model—it is the ability to turn a question into a repeatable workflow.

Source domains have included prices, fundamentals, public regulatory filings, weather and market-microstructure data. Each comes with different coverage and timestamp constraints. Access to a dataset does not make every historical use of it valid.

The public website deliberately omits private infrastructure addresses, credentials, trading endpoints and detailed execution logic.

Open-source validation code ↗

02The research domains

Search the full library ↗

Domain 01