I build research software for economics, and I do research that needs it.

Most quantitative economics runs on code that was never meant to survive the paper it was written for — notebooks that depend on a specific CSV in a specific folder, scrapers that break when a government website changes its markup, estimators reimplemented from scratch by every researcher who needs them. The result is a field where reproducing a result is often harder than producing it. I have spent the last two years building the alternative: installable, typed, tested, documented, citable packages that handle the parts of the pipeline nobody should have to rewrite.

That work started with fedfred, a client for the Federal Reserve’s FRED family of APIs, which I wrote because I needed macroeconomic series for a project and found the existing options unmaintained or incomplete. It now has a few tens of thousands of downloads and a DOI. edgar-sec followed for SEC EDGAR disclosure data. toros is the representation layer those clients materialize into financial objects with enforced semantics rather than untyped tables. Lastly, cultivars is the modeling layer above them, targeting the Bayesian and time-varying VAR methods that Python handles poorly compared to R and MATLAB.

The research side runs in parallel and sets the agenda. My main line is NS-SDN, a neural state-space architecture that decomposes a time series into a time-varying trend plus latent oscillatory components whose amplitude, frequency, and phase are emitted at each step from a recurrent state comprising an adaptive spectral decomposition rather than a fixed Fourier basis. The first paper was published at FLAIRS-39. A second on adaptive spectral emission heads is under review at IEEE. It grew out of a question about the 10-Year Treasury through the 2020 shock that no stationary spectral method could answer cleanly. A related project reframes monetary policy as a control problem, training reinforcement-learning agents to learn a central bank’s reaction function inside both linear TVP-SVAR and nonlinear state-space economies, benchmarked against Hinterlang and Tänzer’s Bundesbank work.

I care about the engineering as much as the econometrics, and I think the separation between them is largely artificial. A model you cannot install is a claim, not a result.

Education

University of Miami, Herbert Business School — Coral Gables, FL B.S.B.A., Quantitative Economics & Finance · Minor in Mathematics

Coursework across calculus, linear algebra, probability theory, mathematical statistics, and econometrics.

Next: Looking for masters by research in computer science and computational science program, with a research focus on state-space methods, spectral representation learning, and reinforcement learning for economic control problems. Specifically interested in programs located in Hong Kong, Singapore, Malaysia and Australia.

Appointments

Research Fellow — Intelligent Computer Systems Research Institute (ICSRI), University of Miami

Founder & Lead Developer — toros-dev, an open-source engineering group building financial data infrastructure

How I work

A few commitments that show up in everything I publish:

Correctness over convenience. Full type annotations, mypy-clean, explicit key-existence checks, specific exception classes, no silent failure. If a response shape is wrong, the library says so rather than returning None.

Sync and async as first-class peers. Not a wrapper bolted on afterward — strict parity, so the async client is never the second-class citizen with missing endpoints.

Reproducibility as a deliverable. OpenSSF Best Practices certification, Codecov coverage gates, conda-forge distribution alongside PyPI, and Zenodo-archived releases with DOIs, so a paper can cite an exact version that will still resolve in ten years.

Design before code. Every package here started as a specification argument — what the public surface should be, what belongs in private internals, where the layer boundaries fall — before any implementation existed. The interesting work is in those decisions, not in the typing.

Technical

Primary: Python 3.13 · PyTorch · NumPy · pandas · Polars · Dask · httpx

Also: Swift · C# / .NET · LaTeX · Git

Domains: time-series econometrics · state-space modeling · Bayesian estimation · reinforcement learning · API and SDK architecture · scientific packaging and distribution

Contact

Email: nsunder724@gmail.com GitHub: nikhilxsunder · toros-dev ORCID: 0009-0007-3323-1760 LinkedIn: nikhil-sunder

Open to research collaboration and graduate program inquiries.