Independent technical research

2026

Wavelet-Domain ENSO Forecasting

Eleven linked documents examine tropical-Pacific sea-surface-temperature forecasting in an exactly invertible complex-wavelet space. They comprise a standalone manuscript, a measurement charter, and nine reports running from system validation to a live-deployment handoff.

Contribution

Designed the forecasting programme, implemented the wavelet and model pipeline, ran controlled evaluations, and wrote the reports and metric charter.

Methods

Dual-tree complex wavelets, Earthformer cuboid attention, CMIP6 forecast pretraining, sequence diffusion, EMOS and ECC-Q calibration, hash-gated sealed testing, moving-block bootstrap governance

Outcome

Threads the research programme together without collapsing its evidence tiers. The record includes one pristine sealed-test win, two controlled studies that narrowed the design space, a probabilistic stack whose validation advantage only partly transferred, and a deployment contract for a prospective live test.

Climate Forecasting / Wavelets / Diffusion Models

Research software · Pre-release

Framework-Free C++/CUDA Chronos and Moirai-MoE

Native implementations of two time-series foundation-model families, including forward and backward kernels, checkpoint loading, and fine-tuning without a tensor-framework runtime.

Pre-release; source not public. Small-variant forward checks demonstrated; broader fidelity and performance tests remain open.

Contribution

Implemented the native model runtime and training kernels, mapped pretrained checkpoints, and built numerical comparisons against reference implementations.

Methods

C++20, CUDA, checkpoint conversion, per-tensor numerical comparisons, explicit fidelity gates, native fine-tuning

Outcome

Forward agreement demonstrated for small model variants. Base-variant checks and performance benchmarks remain unrun; full numerical equivalence is not claimed.

C++20 / CUDA / Numerical Verification

2026

Regime-Adaptive Probabilistic Forecasting

A staged programme studies distributional forecasting of BTCUSDT perpetual futures. It begins with frozen foundation-model embeddings, maps where the series carries learnable signal in a pre-registered scan, and ends with a walk-forward comparison of from-scratch and foundation-model forecasters staged on those verdicts.

Contribution

Built the data and experiment pipeline, implemented and adapted forecasting models, and designed the walk-forward comparisons and attribution controls.

Methods

PyTorch, return and excursion quantile supervision, Chronos-T5 and Moirai-MoE backbones, LoRA fine-tuning, conformalized quantile regression, pre-registered verdict thresholds, purged walk-forward evaluation

Outcome

Establishes that the usable signal is distributional rather than directional and maps where it appears across timeframes and horizons. A small from-scratch encoder defines the quality-versus-cost frontier over the foundation-model lanes, while fine-tuning helps in proportion to the deficiency of each frozen representation.

Probabilistic Forecasting / Foundation Models / PyTorch

2022 – 2026

Wavelet-GAN Bitcoin Forecasting

A multi-year research arc on wavelet-conditioned GANs for short-horizon Bitcoin price-change forecasting, from the original concept paper through to a controlled re-evaluation against matched baselines.

Contribution

Developed the wavelet-conditioned model prototypes and rebuilt their evaluation around matched linear and convolutional baselines.

Methods

Dual-tree complex wavelets, conditional GANs, MATLAB and Python, ridge and TCN baselines, chronological validation

Outcome

Threads four linked reports together: concept, first WaveGAN experiment, refined architecture, and a controlled re-evaluation that ultimately favors simple baselines.

Generative Models / Wavelets / Forecasting

2020 – 2026

Selected Topics in Quant Finance

Six standalone studies cover distinct questions in quantitative finance. They examine FOMC statement sentiment against a 2-year yield proxy, intraday spectral signatures, one-second order-book prediction, change-point detection in SPX realized volatility, cross-sectional intraday reversal, and a public-data replication audit of a real-time equity-ML paper.

Contribution

Designed and implemented six empirical studies, from data preparation and baseline construction to statistical evaluation and technical reporting.

Methods

Event studies, NLP sentiment scoring, signal processing, market microstructure, change-point detection, cross-sectional factor research, replication auditing, bootstrap and significance testing, Python

Outcome

A set of careful, reproducible studies, each an honest result about what does and does not hold in intraday, cross-sectional, and risk-signal data.

Alternative Data / Market Microstructure / Risk Signals

June 2026

Doppler-Domain Micro-Motion Tomography

A reproduction, critical analysis, and benchmark of single-SLC Doppler-domain micro-motion tomography: the published method rebuilt as an extended CUDA pipeline, then tested against synthetic ground truth and falsification-oriented null tests.

Contribution

Reconstructed the published SAR method in C++/CUDA and designed synthetic ground-truth, null, and robustness tests.

Methods

C++, CUDA, OpenMP, TerraSAR-X SSC processing, synthetic ground truth, null tests, 48-hour benchmark suite

Outcome

Documents five signal-level structural results omitted by the source papers and implements the corrections they imply. The front end recovers injected micro-motion, while tomographic localization does not clear its own pre-registered shuffle-pairs and sign-flip floors.

SAR Tomography / C++ / CUDA / Research