00 — finance
The boring half of quant finance, done well. Tabular DL only when it earns its place.
For hedge funds, asset managers, and portfolio teams who want code their PMs can read on a Friday.
baseline
GBDT
before any DL touches your data
typical scope
6–12 wks
read, fix, hand back
reporting
tearsheets
no 80-slide appendix
01 — what we do
Concrete deliverables. Named tools. Honest constraints.
01
Portfolio construction with ML overlays.
Mean-variance and risk-parity baselines first; an ML overlay on top only when the eval shows it pays for its own variance.
02
Factor research, end to end.
Hypothesis to backtest to capacity analysis to sizing. Documented enough that your PM can argue it down or fund it up.
03
Risk infra.
Exposures, scenario PnL, factor decomposition, and stress runs. Numbers refresh on a schedule, not on request.
04
Backtest tooling.
A backtester your team trusts: point-in-time data, realistic costs, and a path from a notebook idea to a reproducible run in under a day.
05
Tabular DL where it earns its place.
Gradient-boosted trees as the baseline. We move to TabNet, FT-Transformer, or sequence models only when a controlled eval beats GBDT on the metric that matters.
06
Reporting your IC will read.
Tearsheets that say what changed, why, and what the next decision is. No 80-slide appendix.
02 — how we engage
We will not rewrite a working backtester to look like ours. We will fix the parts that lie to you.
Typical scope is 6 to 12 weeks. We start by reading the research stack and one quarter of historical PnL before proposing changes.
Code is delivered in your repo, in your stack. We leave with a handover doc, not a key dependency.
03 — example shape
an engagement shape
Anonymised. Plausible. The shape of a real engagement.
A multi-strategy fund had a tree of forty alpha researchers and a backtester that disagreed with production by 90 bps a year.
The gap was point-in-time data and three subtle look-ahead bugs in the join layer.
We rebuilt the join, ran a parallel reconciliation against six months of live fills, and shipped a new backtest harness over 7 weeks.
The discrepancy fell to 6 bps, all of it cleanly attributable to costs.
Two of the forty alphas were retired the week after — they had been backtest artefacts the whole time.
The desk kept the harness. We left.
04 — contact
Tell us where the work is.
One sentence is enough. We reply within two business days. If we are a bad fit we will say so.