The implementations behind the research: data pipelines, model architectures, experiment harnesses, and evaluation tools.
Each study has a companion page describing what its implementation is built out of, how it runs, and what keeps its results reproducible; a few projects appear here without a companion study because the software is the work. Most of it is not publicly released, and those pages say what a release would require; where code or data has been published, the links are on the page.
Numerical Verification · GPU Computing · Time-Series Foundation Models
Two pretrained time-series foundation models reimplemented as framework-free C++/CUDA with hand-written forward and backward passes, and a verification harness built around what a reimplementation can honestly claim to have established.
Signal Processing · Machine Learning · Climate
The substrate the whole series runs on: an exactly invertible complex-wavelet representation, a coefficient-native adaptation of an external pretrained backbone with audited checkpoint surgery, and a data boundary that fails closed on the sealed period.
Signal Processing · Machine Learning · Climate
The Phase A ablation harness: nine controlled adaptation families over 335 observed windows, a persistence-residual objective, an explicit freeze policy, and comparison writers that report amplitude damping and residual cancellation alongside error.
Signal Processing · Machine Learning · Climate
Masked-coefficient pretraining over four historical SST reconstructions, a multistore ingestion contract with per-store normalization and validity masks, a coefficient-energy audit, and the paired moving-block bootstrap the rest of the series is governed by.
Signal Processing · Machine Learning · Climate
Simulation-scale forecast pretraining: public CMIP6 catalogue discovery, fourteen simulation stores behind a strict ingestion contract, a 30,000-step exactly resumable CUDA loop with held-out-model selection, and the hash-gated protocol behind the series’ one sealed-test read.
Signal Processing · Machine Learning · Climate
A second input block for upper-ocean heat content, added as a controlled experiment: matched simulation and observed OHC stores, a 74-channel dual-predictor model, and a zeroed-input twin that isolates the predictor from everything else about the run.
Signal Processing · Machine Learning · Climate
Three probabilistic heads around a frozen deterministic centre — a per-coefficient Gaussian, a shared-latent-noise ensemble, and a conditional diffusion denoiser over correction fields — with a common ensemble-evaluation layer that reports field and regional calibration separately.
Signal Processing · Machine Learning · Climate
A six-rung calibration ladder in the regional-index projection space: lead-smooth variance fitting, quantile remapping, leave-one-year-out protocol, state-conditional and blended variants, and a lifting step that returns calibrated index members to coherent fields.
Signal Processing · Machine Learning · Climate
A joint fourteen-lead diffusion head with spatio-temporal blocks, band-weighted objective and correlated noise, plus the trajectory metrics and the protocol amendment authorizing the series’ second and final sealed-test read.
Signal Processing · Machine Learning · Climate
Five preprocessing pipelines that differ only in the anomaly reduction, each retrained from scratch: fixed legacy and WMO climatologies, a trailing block, a relative-SST frame and a linear detrend — plus the record extension and the frozen deployment contract.
Signal Processing · Machine Learning · Climate
The measurement layer the other ten documents are scored by: persistent metric identifiers, the house paired moving-block bootstrap and its resolvability standard, protocol gates, an append-only validation ledger, and the audits that check the record against itself.
Machine Learning · Probabilistic Forecasting · Financial Markets
A GPU-first walk-forward comparison package: a compact multi-resolution encoder against frozen and LoRA-fine-tuned foundation-model lanes, monotone quantile heads, conformal recalibration, triggered online adaptation, and a study matrix that emits the manuscript’s tables directly.
Signal Processing · Machine Learning · Climate
The publication build: the adopted deterministic forecaster and a 32-member residual-diffusion ensemble in the relative-SST frame, with an audited event-case gallery, a finalization step, and an eleven-script editorial revision suite adding statistical baselines and sensitivity analyses.
Signal Processing · Time–Frequency Analysis · Financial Markets
A symmetric decomposition pipeline: matched-session SPY and BTC returns, parallel STFT and discrete-wavelet band shares, date-block bootstrap comparisons, and a SHA-256 manifest tying every published number back to its inputs.
Machine Learning · Probabilistic Forecasting · Financial Markets
A deliberately small pre-registered scanner: causal per-timeframe features, leak-safe multi-resolution alignment, five target families, small GPU probes, and verdict thresholds fixed before the scan that a cell must clear on validation and test alike.
Machine Learning · Probabilistic Forecasting · Financial Markets
A GPU head-research harness over cached frozen foundation-model embeddings: a memory-mapped embedding store, purged and embargoed splits measured in time, monotone quantile heads, a collapse detector, permutation controls, and cost-aware evaluation.
Statistical Inference · Time Series · Financial Markets
A resumable minute-bar ingestion and factor-research pipeline: point-in-time universe construction, multi-horizon decile sorts, block-bootstrap intervals, decay fitting, cost overlays, and a publication gate that fails closed on demo output.
Natural Language Processing · Statistical Inference · Financial Markets
FOMC statement scraping and cleaning, sentence-level FinBERT scoring at a pinned model revision, event-window alignment to the release minute, and HAC-robust regression with window, cycle and outlier sensitivity computed on every run.
Change-Point Detection · Time Series · Financial Markets
An installable package running HMM, CUSUM, PELT and a fixed-threshold baseline through one evaluation protocol, with pinned event matching, ablation grids, data-vintage evidence, and a change-point export that other studies consume.
Signal Processing · Forecasting · Financial Markets
A config-driven MATLAB research prototype in twelve numbered stages: cleaning, chronological splits, windowing, wavelet feature extraction, four baselines and three residual temporal-convolutional models, multi-seed training, return-domain evaluation, and a reproducible report bundle.
Machine Learning · Replication · Financial Markets
An expanding-window annual machine-learning pipeline built for auditing rather than for a headline: schema contracts on every stage, recorded source vintages, 28 registered ablations, a gap decomposition, and a production gate that fails closed on smoke paths.
Machine Learning · Forecasting · Financial Markets
Order-book reconstruction from LOBSTER message and book files, one-second-ahead labelling, a chronological logistic-regression evaluation sliced by time of day and spread, and a post-cost overlay that keeps AUC from being read as profitability.