loginil/Kronos-small-btc-15m-dir-zeroshot

Path G — DirectionKronos with frozen NeoQuasar/Kronos-small backbone. Val acc 50.45% (essentially random).

Architecture

• Backbone: frozen Kronos-small (24.7M, no grads)
• Head: mean-pool last 90 hidden states → Linear(512→256)→GELU→Dropout→Linear(256→2)
• 131K trainable params (200× less than backbone)
• Output: softmax([P(DOWN), P(UP)])

Usage

from model.direction_kronos import DirectionKronos
import torch.nn.functional as F
import numpy as np

model = DirectionKronos.load(
    head_path="loginil/Kronos-small-btc-15m-dir-zeroshot",  # or pathe
    backbone_path="NeoQuasar/Kronos-small",                 # for zeroshot
    # backbone_path="loginil/Kronos-small-btc-15m-v3",     # for pathe
    tokenizer_path="NeoQuasar/Kronos-Tokenizer-base",
)
# x: (B, 90, 6) instance-normalized OHLCV+amt
# stamp: (B, 90, 5) time features
probs = model.predict_proba(x, stamp)  # (B, 2) [P(DOWN), P(UP)]

Result

No edge found. Val accuracy is statistically indistinguishable from random (50%). Walk-forward on fresh 2026-03 data needed to confirm.