imbue-bit/ns-ntk

★ 14PythonAudience · researcherComplexity · 5/5Setup · hard

Mindmap

mindmap
  root((ns-ntk))
    Research topic
      Non-Stationary NTK
      Continuous distribution shift
      Quantitative finance
    Components
      PVR optimizer
      EAT optimizer
      Exponential Aggregation sampler
      Temporal Transformer
    Audience
      ML researchers
      Quant finance researchers
    Limitations
      No setup docs
      Paper not public
      Research use only

mindmap root((ns-ntk)) Research topic Non-Stationary NTK Continuous distribution shift Quantitative finance Components PVR optimizer EAT optimizer Exponential Aggregation sampler Temporal Transformer Audience ML researchers Quant finance researchers Limitations No setup docs Paper not public Research use only

Click or tap to explore — scroll the page freely

Things people build with this

USE CASE 1

Reproduce experiments from the NS-NTK paper on neural network behavior under continuous market distribution shift

USE CASE 2

Use the Temporal Transformer implementation as a reference for building time-series neural networks for financial data

USE CASE 3

Study the PVR and EAT optimizers as examples of training algorithms designed for non-stationary environments

Tech stack

Python

Getting it running

Difficulty · hard Time to first run · 1day+

No setup instructions provided, requires reading the linked paper (not publicly hosted) to understand how to use the components.

No license information was mentioned in the explanation.

In plain English

This repository is the official code release for an academic paper titled "Deep Learning under Continuous Distribution Shift: The Non-Stationary NTK and Spectral Tracking SDE for Quantitative Finance." The paper appears to have been posted in May 2026, though it is only linked as a local PDF and is not publicly accessible through this README. The subject matter sits at the intersection of machine learning theory and financial markets. The core idea concerns how neural networks behave when the data they were trained on keeps changing over time, which is a common challenge in finance where market conditions shift continuously. The repository name "NS-NTK" refers to the Non-Stationary Neural Tangent Kernel, a mathematical framework for analyzing how neural networks learn. According to the README, the code includes several components: PVR and EAT optimizers (training algorithms), Exponential Aggregation samplers (a method for combining data or model outputs), and Temporal Transformers (a type of neural network designed for time-series data). No explanation of how to install or run any of these is provided in the README. The documentation is very sparse. Beyond the component list and a reference table image, there are no setup instructions, usage examples, or explanations of what the tools actually do in practice. Anyone who wants to use this code would need to read the linked paper for context, and that paper is not publicly hosted through this repository. This is a research code release aimed at other machine learning researchers, not a general-use library.

Copy-paste prompts

Prompt 1

I have ns-ntk cloned and have read the NS-NTK paper. Walk me through running the PVR optimizer on a simple financial time-series dataset to reproduce the paper's basic experiment.

Prompt 2

How does the Temporal Transformer in ns-ntk differ from a standard transformer? I want to understand what architectural changes make it suited for non-stationary financial data.

Prompt 3

I want to use the Exponential Aggregation sampler from ns-ntk in my own PyTorch training loop. Show me how to import and configure it for a custom dataset.

Prompt 4

What is the Non-Stationary Neural Tangent Kernel concept that ns-ntk is based on, and how does it change how we should think about training neural networks on shifting data?

Open on GitHub → Explain another repo

← imbue-bit on gitmyhub — every repo by this author, as a profile.

Verify against the repo before relying on details.