Tycheon: Calibrated Uncertainty, Risk Translation and Leakage-Guarded Evaluation for Financial Time-Series Foundation Models
Year: 2026
Type: preprint
Abstract: On a small synthetic benchmark (3 series, 108 pooled origins), no model, including zero-shot Kronos-mini, was statistically distinguishable from the random walk, and Kronos-mini's nominal 90% interval contained 57.4% of outcomes. A single manual check on real daily equity data showed two-sided conformal calibration, fitted on 60 recent origins, reducing a 5-day 95% VaR from 5.38% to 3.93%. This led to the rule that calibration may only widen the tails used for risk figures. Multi-asset portfolio risk is always labelled uncalibrated, because dependence between assets is assumed. Neither result is evidence of forecasting skill in real markets. Software: https://doi.org/10.5281/zenodo.24041298Code: https://github.com/anilatambharii/tycheonDemo (synthetic data): https://huggingface.co/spaces/AmbhariiLabs/tycheon-demo For research and risk analytics. Not investment advice.
Author: Prasad, Anil
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7225460718