sohail-kustagi/Qwen3-14B-xauusd-v1

Qwen3-14B — XAUUSD Multi-Agent Trading Synthesizer (v1)

Model Description

This model is a LoRA fine-tune of Qwen3-14B, specialized for synthesizing multi-agent market analysis data for XAU/USD (Gold) algorithmic trading.

It was trained on a proprietary dataset of 5,044 structured reasoning chains generated by a live Bloomberg-style multi-agent trading system, including:

• Confidence Scoring — Evaluating technical confluence across M5/M15/H1 timeframes using EMA crossovers, RSI, ADX, and ATR-based volatility metrics.
• Macro Sentiment Synthesis — Distilling live RSS macroeconomic headlines (Fed policy, DXY correlation, Treasury yields) into a scalar sentiment score (−1.0 to +1.0) for XAU/USD bias.
• Historian RAG Matching — Scoring structural similarity between live price patterns and a vector database of historical XAUUSD sequences.
• Risk Management — Generating ATR-dynamic Stop Loss and Take Profit levels. Static pip distances are explicitly forbidden in the training data.
• Live Copilot Advice — Producing in-trade policy recommendations (HOLD / ADJUST_POLICY / CLOSE) for active XAUUSD positions.

Intended Use

This model is designed as the inference backbone for a real-time XAUUSD trading terminal, replacing generic LLM API calls with a domain-specialized model that understands institutional trading vocabulary natively.

Training Details

Parameter	Value
Base Model	Qwen/Qwen3-14B
Fine-tuning Method	LoRA (Low-Rank Adaptation)
Training Method	SFT (Supervised Fine-Tuning)
LoRA Rank	16
LoRA Alpha	32
LoRA Trainable Modules	all-linear
Epochs	1
Batch Size	8
Learning Rate	2e-4
LR Scheduler	Cosine
Warmup Ratio	0.03
Sequence Packing	True
Max Sequence Length	40,960
Training Tokens	~3.2M
Hardware	H100-80GB SXM (Together AI)
Training Time	~10 minutes

Dataset Composition

Agent	Records	Description
CONFIDENCE	3,569	Market state → BUY/SELL/WAIT + SL/TP + confidence
NEWSHOUND	840	Macro headlines → sentiment scalar
HISTORIAN	562	Price pattern → RAG similarity score
COPILOT	65	Live trade state → management policy
Total	5,044

System Architecture

The model operates within a 5-agent asynchronous pipeline:

WebSocket Feed (FXOpen)
    ├── RegimeDetective (HMM: TRENDING/RANGING/VOLATILE)
    ├── TrendAnalyzer (Multi-Timeframe EMA/RSI/ADX)
    ├── VolumeDetector (Surge detection)
    ├── HistorianRAG (Vector DB similarity)
    └── NewsHound (RSS → FinBERT-style scoring)
            ↓
    [This Model] — Confidence Meter + Copilot
            ↓
    OrderGuardian (ATR-based SL enforcement)

Limitations

• Trained on XAU/USD data only. Not suitable for other trading pairs.
• Training used train_on_inputs=True due to general-format dataset detection. A future version will use properly structured chat-format data for improved masking.
• 1 epoch training — a second fine-tuning round with Desk Manager data is planned.

Future Work

• Round 2 Fine-Tune: Adding "Desk Manager" agent data — a unified personal trading advisor that synthesizes all agent outputs with the trader's personal performance profile (win rate, drawdown, streak state) to produce tailored entry/exit setups.
• RLHF Integration: Outcome-labeled records (trade result vs. model advice) will enable preference-based fine-tuning for reward-signal alignment.