menes000/chronos-2-xauusd-3m-lora

Chronos-2 LoRA for gold (XAUUSD), 3-minute bars

A LoRA adapter for amazon/chronos-2, fine-tuned on 10 years of XAUUSD (gold/USD) 3-minute closing prices to forecast the next 1 to 3 hours.

Usage

import torch
from chronos import Chronos2Pipeline

pipe = Chronos2Pipeline.from_pretrained("menes000/chronos-2-xauusd-3m-lora", device_map="cpu")

# context: the last 2048 3-minute closing prices (market-closed periods removed, bars consecutive)
context = torch.tensor(closes_3m[-2048:], dtype=torch.float32)
quantiles, mean = pipe.predict_quantiles(
    [context], prediction_length=60, quantile_levels=[0.1, 0.25, 0.5, 0.75, 0.9]
)
# quantiles[0]: [1, 60, 5] -> step 20 = 1 hour ahead, step 60 = 3 hours ahead

Requires chronos-forecasting>=2.3 and peft. The base model is downloaded automatically.

Training

Data	XAUUSD 1-minute OHLC, resampled to 3-minute closes
Train	2013-01 → 2022-12 (~1.18M 3-minute bars)
Augmentation	3 phase-shifted 3-minute series (bar boundaries at :00/:03, :01/:04, :02/:05)
Validation	512 random windows from 2023
Test	2024 and 2025-01 → 2026-08, never seen during training
Context / horizon	2048 bars (~4 trading days) / 60 bars (3 hours)
Method	LoRA r=8, alpha=16 on self-attention q/k/v/o and the output layer (~1.2M trainable params)
Optimisation	1000 steps, batch 64, lr 1e-5 linear decay, AdamW

Validation loss (2023): 2.2800 (base) → 2.2202. Full curve in training_log.csv.

Results: LoRA vs base Chronos-2

The adapter was evaluated on two test periods that were never used for training or model selection (2024 and 2025-01 → 2026-08), 600 non-overlapping windows each. Both models received exactly the same inputs. The metric is the quantile loss over the 0.1 / 0.25 / 0.5 / 0.75 / 0.9 forecasts at the end of the horizon (lower is better).

Period	Horizon	Base Chronos-2	LoRA	Error reduction (95% CI)	LoRA better in
2024	1 hour	1.177	1.161	−1.3% [0.2%, 2.4%]	63% of windows
2024	3 hours	2.134	2.088	−2.1% [1.1%, 3.3%]	63% of windows
2025–26	1 hour	3.285	3.211	−2.2% [0.6%, 4.2%]	61% of windows
2025–26	3 hours	5.327	5.134	−3.6% [2.2%, 5.1%]	62% of windows

• The improvement is statistically significant in all four cases (the 95% interval excludes zero).
• It holds in both test periods, including 2025–26, when gold traded at 2,600–5,600 USD with much larger dollar moves than anything in the training data (which peaked near 2,070 USD).
• The gain grows with the horizon: larger at 3 hours than at 1 hour.
• Validation loss on 2023 fell from 2.2800 to 2.2202 (−2.6%) and decreased steadily without overfitting.

Loss values are in USD per ounce, so they are larger in 2025–26 simply because gold moved more.

Trading

Despite this, a trading strategy based on its forecasts is not profitable once trading costs (0.30 USD/oz per trade) are taken into account.

Not financial advice.