---
title: Kronos Market Forecast
emoji: 📈
colorFrom: purple
colorTo: indigo
sdk: docker
app_port: 7860
pinned: false
license: mit
---

# Kronos Market Forecast

A small web app that turns the open-source [Kronos](https://github.com/shiyu-coder/Kronos)
financial foundation model into an interactive, multi-asset forecasting tool — plus a
surface-roughness research lab for calibrating and fine-tuning it.

> ⚠️ **Research / educational demo — not financial advice.** Forecasts are probabilistic
> samples from a model pre-trained largely on crypto K-lines, and are frequently wrong.
> The leverage optimizer is a mechanical translation of a forecast into trade structure,
> not a recommendation. Leverage can lose your entire margin (and more). Do not trade on this.

## What's inside

| Component | Path | What it does |
|---|---|---|
| **Chart UI** | `crypto_ui/` (`/`) | Search any asset (crypto via Binance, stocks/ETFs/FX/indices/commodities via Yahoo), forecast it with Kronos-small/base, see ghost-candle predictions + a p10–p90 uncertainty band. |
| **Portfolio Analyzer** | `crypto_ui/` (`/analyzer`) | Forecast one ticker across a matrix of intervals × horizons × both models; color-graded heatmaps, model-agreement consensus, save to JSON + standalone HTML report. |
| **Leverage Trade Optimizer** | `crypto_ui/app.py` (`/api/optimize`) | Turns a forecast cell into Conservative/Balanced/Aggressive trade setups (entry, stop, take-profits, leverage sized so the stop sits inside liquidation) with full reasoning. Suppresses setups when the two models disagree on direction. |
| **Surface-roughness lab** | `roughness_lab/` | Treats price as a measured surface profile (ISO-16610 Gaussian waviness/roughness split; Ra/Rq/Rz/RSm/Rsk/Rku/Wa). Calibrates Kronos sampling settings so forecast *texture* matches real market texture. |
| **GPU fine-tune pipeline** | `roughness_lab/gpu_finetune/` | Roughness-aware fine-tuning: per-epoch checkpoints selected by texture realism, ready to run on a GPU box. |

## Run locally

```bash
python -m venv .venv
.venv/Scripts/python -m pip install -r crypto_ui/requirements.txt   # Windows
# source .venv/bin/activate && pip install -r crypto_ui/requirements.txt  # macOS/Linux
python crypto_ui/app.py
# open http://127.0.0.1:8765  (analyzer at /analyzer)
```

Models (`NeoQuasar/Kronos-small`, `NeoQuasar/Kronos-base`, `NeoQuasar/Kronos-Tokenizer-base`)
download automatically from the Hugging Face Hub on first use. CPU works; a forecast takes
~15 s (small) to a few minutes (base). The app imports the Kronos `model/` package — when
deployed, that package is vendored alongside the app (see deploy notes).

## Credits & license

Built on [shiyu-coder/Kronos](https://github.com/shiyu-coder/Kronos) (MIT). Kronos:
*A Foundation Model for the Language of Financial Markets*, Shi et al., AAAI 2026
([arXiv:2508.02739](https://arxiv.org/abs/2508.02739)). This project is MIT-licensed; the
vendored `model/` package retains its upstream MIT license and copyright.