Technical documentation on TradeStance's pricing models, data pipeline, and analytical framework.
P_L = (P_F + C_L + C_S + C_I) x (1 + T) + D + F
TradeStance employs a hierarchical waterfall strategy to find the most accurate price for any commodity:
Direct feeds from major global commodity exchanges (LME, ICE, CBOT) for real-time spot and futures pricing.
Official government commodity boards and statistical offices providing reference prices and export data.
UN, IMF, and FAO datasets providing trade statistics, commodity price indices, and food price monitoring data.
Licensed data providers supplying financial market data, economic indicators, and commodity-specific pricing.
Structured extraction from exchange-specific websites and regional commodity boards with multi-source cross-validation.
Proprietary AI model cross-references multiple sources to establish consensus pricing with statistical outlier rejection.
Every data point passes through a multi-layer validation gate before entering the system:
Rejects prices below physically plausible minimums based on commodity-specific thresholds.
Flags prices exceeding historical percentile bounds as potential errors or manipulation.
Purges entries older than configurable time windows from active price surfaces.
Dynamic scoring of each data source based on accuracy history, freshness, and consistency.
Statistical analysis comparing new prices against rolling historical windows for deviation detection.
Multi-source agreement requirement: multiple sources must converge within tolerance for high-confidence pricing.
Architecture: Dual-stream Integrated LSTM with Attention
|-- Temporal Stream
| |-- Input: Price sequences with configurable lookback
| |-- Features: log_return, volatility, moving average convergence
| +-- LSTM with multi-head attention mechanism
|-- Semantic Stream
| |-- Input: News sentiment embeddings
| +-- Dense layers for feature compression
+-- Fusion Layer
|-- Concatenation of temporal and semantic outputs
|-- Dense layers with dropout regularization
+-- Output: Forward price forecast with confidence intervalsThe TradeStance news pipeline generates market intelligence at regular intervals without human intervention:
Scheduled Trigger
|
|-- 1. Ingest Raw Data
| |-- Commodity exchange prices
| |-- Government reference prices
| +-- Spot prices from major exchanges
|
|-- 2. News Aggregation
| |-- Global trade headline collection
| |-- Commodity-specific coverage filtering
| +-- Deduplication and relevance scoring
|
|-- 3. AI Processing
| |-- Generate concise Trade Briefs
| |-- Extract: country, commodity, sentiment
| +-- Professional financial reporting tone
|
|-- 4. Sentiment Scoring
| |-- NLP classification: BULLISH / NEUTRAL / BEARISH
| +-- Confidence-weighted impact scores
|
+-- 5. Storage & Distribution
|-- Database persistence
|-- Real-time broadcast to connected clients
+-- System health report generation