THE HARD PART
What could not be hand-waved
The temptation in this space is to claim a state-of-the-art result. Most such claims come from evaluation that leaks — a random split on time-ordered data will beat an honest one every time.
ENGINEERING CASE · Open source · C++17 · Apache-2.0
Purged GBDT Engine (TideGBM)
Our own gradient-boosting library for time-indexed and panel data, published openly on GitHub and PyPI — with purged validation and a leak lint built in.
The task: Standard gradient boosting treats rows as exchangeable. Market data is not: it has an arrow of time and an entity structure, and evaluating it without purging produces backtests that look excellent and lose money.
DELIVERY
What was built
- Column-major C++17 histogram core with OpenMP, and optional CUDA histograms
- Temporal ordered residual boosting — causal leaf residuals that respect time order
- Purge and embargo on walk-forward splits with horizon-aware gaps
- Strict leak lint that refuses features which peek into the future
- Within-date pairwise ranking objective for cross-sectional information coefficient
- scikit-learn-compatible Python API, ONNX export, continuous integration
EVIDENCE
Engineering measurements
Apache-2.0open licence
PyPIpublished as tidegbm
C++17core with OpenMP, optional CUDA
Publicsource at github.com/Vizanix-Team
Figures describe software behaviour and come from the running system. Client identity, strategy parameters and financial results stay under NDA; the cover is a schematic, not a client screenshot.
THE RESPONSE
How the risk was addressed
Purging and embargo are the default rather than an option, and strict mode refuses to fit on features that fail the look-ahead lint. The stated goal in the documentation is a stronger gradient booster for tabular time and panel problems under honest, purged evaluation — explicitly not a claim of beating neural networks. The code is public, so the claim is checkable.
START A PROJECT
Tell us what you want to build
A short description and one contact are enough. Do not send exchange keys, passwords or sensitive data.
Or contact us directly · @vx_ceo