RESEARCH THAT MATTERS.
Duon Labs conducts AI research for forecasting chaotic systems, developing frameworks that challenge conventional assumptions while maintaining scientific rigor in our open methodologies.
RESEARCH DOMAINS
DISTRIBUTION MODELING
A new paradigm for representing uncertainty.
SICK BIDE introduces Binary Implicit Distribution Encoding, a neural architecture that models probability distributions via binary representations, paired with GPU-optimized softmax integral computation. Enables exact, assumption-free inference across high-dimensional, continuous and categorical domains.
SCALING LAWS FOR PREDICTABILITY
Inspired by LLM scaling trends.
Apogée investigates whether increased compute, data, and model size improve crypto market forecasting. We quantify how many bits of future price action can be inferred from historical data, and explore the limits of time-series predictability using deep learning.
INFORMATION-THEORETIC TIME SERIES ANALYSIS
Time-series forecasting as information extraction.
Our research leverages extropy as a measure of predictive information, focusing on signal density, volatility clustering, and long-range temporal dependencies.
GPU-ACCELERATED FORECASTING INFRASTRUCTURE
Fast, interpretable, and scalable.
We build production-grade forecasting stacks. Our tooling includes custom GPU kernels, efficient Monte Carlo simulation, and integrated model evaluation pipelines, enabling deployment with scientific fidelity.
RESEARCH PAPERS
SICK BIDE: SOFTMAX INTEGRAL COMPUTE KERNEL + BINARY IMPLICIT DISTRIBUTION ENCODING
A foundational framework for modeling probability distributions directly from binary representations of data. Combines a lightweight MLP architecture with GPU-accelerated softmax integral computation. Enables efficient, exact Monte Carlo inference without approximations. Applicable to time-series modeling, LLM output layers, and beyond.
APOGÉE: SCALING LAWS FOR CRYPTO MARKET FORECASTING
Experimental framework to quantify the limits of predictability in crypto markets. Measures extractable information using deep learning and scaling experiments across model size, dataset scope, and compute. Benchmarks extropy gain per asset and timeframe. Designed as a reproducible, open benchmark for time-series forecasting in finance.
RESEARCH PRINCIPLES
OPEN-SOURCE ETHOS
- Core libraries and tools released under open licenses
- Reproducible pipelines and training scripts
- Modular design for community reuse and benchmarking
CALIBRATED EVALUATION
- Probabilistic scoring metrics (extropy, CRPS, proper scoring rules)
- Multi-horizon evaluation for short- and long-term reliability
- Time-series aware validation with no look-ahead bias
SCIENTIFIC ENGINEERING
- Built for integration: APIs, SDKs, and CLI tools
- GPU-accelerated by default for scalability
- Infrastructure tested under real-world constraints
RESEARCH COLLABORATION
We partner with researchers and institutions focused on transparent, scientific forecasting. If you work on time-series AI, scaling laws, or simulation, reach out.