Shadowell/Kairos-base-crypto

Fine-tuned Kronos-base on BTC/USDT + ETH/USDT 1-min K-lines (2024-01 ~ 2026-04) using Kairos. Architecture = Kronos + exogenous bypass channel (32-d) + quantile return head.

This run keeps the original tokenizer NeoQuasar/Kronos-Tokenizer-base, matching the original Kairos-small-crypto training flow. Training data comes from the public Binance Vision spot mirror, so the 5 crypto-native exogenous features (funding_rate / funding_rate_z / oi_change / basis / btc_dominance) remain padded to zero; the other 27 dimensions are real.

Results on test set (2026-01-01 04:16:00 ~ 2026-04-16 23:30:00, 304,710 1-min bars)

horizon	model	hit_rate	rank_ic	ICIR
h1	baseline	50.78%	+0.025	+0.630
h1	finetuned	50.37%	+0.011	+0.051
h5	baseline	51.61%	+0.029	+0.385
h5	finetuned	50.95%	+0.029	+0.351
h30	baseline	52.49%	+0.055	+0.325
h30	finetuned	52.92%	+0.076	+0.484

Baseline = original Kronos-base weights + randomly initialised exog / return head. Tokenizer stays on the official NeoQuasar/Kronos-Tokenizer-base, matching the original crypto predictor flow. Training stopped at epoch 4; best val_ce = 2.4842.

Usage

from kairos import KronosTokenizer, KronosWithExogenous
tok = KronosTokenizer.from_pretrained("NeoQuasar/Kronos-Tokenizer-base")
model = KronosWithExogenous.from_pretrained("Shadowell/Kairos-base-crypto")

Training config (preset crypto-1min)

• lookback 256 min, predict 30 min
• batch 24, OneCycleLR, early-stop patience 3
• progressive unfreeze: only last transformer block + exog bypass + return head
• tokenizer source = NeoQuasar/Kronos-Tokenizer-base
• 32-d EXOG = 24 common + 8 crypto-market features

Training recipe

Full command log, backtest commands, pitfalls and the reproduction checklist are in docs/CRYPTO_BTC_ETH_RUN.md.