About the Project
SILTHRA is a research-driven financial console for intelligent dialogue on Russian equities. The language layer is powered by GigaChat; the server backend is built on FastAPI running on a Linux VPS, with dedicated modules for data ingestion, calculations, and generating well-structured market scenarios.
We currently utilize MOEX ISS (HTTP/JSON) with pagination handling and field normalization. In parallel, we are preparing integration with the MOEX Real-Time API (Non-Display) under a commercial agreement — enabling low-latency and extended data feeds. The system manages timeouts and quotas, caches results, and gracefully degrades in case of outages.
How It Works
- Issuer recognition. We normalize text (RU/EN, grammatical cases), map to SECID/shortname/name and brand aliases, then accurately resolve to the MOEX ticker (TQBR).
- Data acquisition. Quotes, dividends, issuer profile, and daily history are pulled via ISS with cursor pagination, accounting for gaps and non-trading days.
- Technical module. Long RSI, its SMA, and Bollinger Bands on RSI with custom parameters; behavioral patterns (crossovers, retests, breakouts/returns to range) → phase assessment and key targets.
- Explanations. GigaChat produces concise buy-side style narratives without disclosing exact indicator values: trend, key zones, probable targets, and risks.
- Interactive prompts. Buttons like “Trend & Targets,” “Quotes,” “Dividends,” and “History” trigger backend actions without unnecessary typing.
Architecture
VPS→FastAPI→MOEX-ISS/Real-Time API→TA modules→GigaChat. All clients have retries and time limits; errors are logged without leaking technical details to the frontend.- Russian text normalization (declensions), an auto-updated issuer directory, and a fast ticker resolver.
- Technical analysis runs in a thread pool, with “soft” time limits, cache warming, and safe pattern matching.
Research & Backtesting
We identify stable behavioral patterns on long-term data, validate them through backtests (including TradingView/Pine scripts for cross-checking ideas), and turn them into risk-controlled scenarios. Historical data spans back to IPO where available, with adjustments for gaps and corporate actions.
- RSI → SMA → BB (on RSI). A framework for detecting phase shifts, confirming trends, and defining working ranges.
- RSI support/resistance. Dynamic RSI levels are projected into price targets via reverse calculations against the current range.
- Scenarios. “Cross/Retest,” “Breakout/Return beyond BB,” and “Consolidation near the mean” — tracked with hit/miss and drawdown metrics on rolling windows.
Disclaimer: All materials are for research purposes only and do not constitute investment advice.
Data & Agreements
Currently — MOEX ISS (HTTP/JSON). In progress — Non-Display agreement for MOEX Real-Time API (low latency, extended fields). Data is used exclusively for internal computations and is never redistributed in raw form.
Roadmap
- Real-time MOEX streams, including order book and market event features.
- Expanded feature library (momentum, reversion, volatility regimes, consolidation patterns).
- Hybrid scenario ranking (LLM + structured feature scoring).
- Dedicated GPU infrastructure and private LLMs trained on proprietary domain knowledge bases.
- Assistants for financial reporting, taxation, corporate actions, and legal compliance.
Processing Pipeline
Simplified process flow: text normalization → issuer recognition → historical data loading (from IPO) → feature calculation (RSI/SMA/BB on RSI) → scenario & target generation → buy-side style explanation via GigaChat.