About SILTHRA
SILTHRA is an ongoing effort to build a professional intelligence layer for listed markets. We blend quantitative research, narrative, and a conversational interface to support faster, clearer decisions.
- Active build — iterative releases, controlled surface area.
- Server‑side computation; the interface presents conclusions, not raw noise.
- Implementation details are deliberately private.
Data Sourcing
We are implementing resilient collectors from official endpoints with normalization and outage handling. Historical ranges are merged with recent prints when available.
- Unified time series across history and live checks.
- Issuer/ticker resolution for aliases and languages.
- Backoff, caching, quota‑aware polling.
News Flow (Design)
A pipeline under design tags headlines by issuer and routes only relevant context to the language layer within seconds.
- Symbol‑aware parsing and relevance scoring.
- Time‑ordered digest with minimal duplication.
- Latency‑first delivery to the interface.
Strategy & Tuning (Work in Progress)
We’re building zone‑based momentum engines designed to capture directional shifts with disciplined exits. Engines use a fast signal plus a slower context gate. Entries require alignment; exits may occur when evidence flips. Horizons and sensitivities are tunable per asset.
- Zone logic: Adaptive bands with explicit cross/retest handling. No repainting — history is replayed prefix‑by‑prefix.
- Targets vs. hold: One path projects momentum into price‑space levels; another uses target‑free, bar‑based holds. Signal generation and execution rules are kept separate.
- Execution realism: Commission, slippage, and entry delay are modeled; stops may be time‑ or excursion‑based.
Tuning is under active R&D. We are experimenting with adaptive search methods and multi‑horizon walk‑forward evaluation to reduce leakage and overfitting.
- Seeded exploration followed by focused refinement.
- Supports asymmetric sensitivities across signals and context gates.
- Parallel evaluation with immutable audit trails.
- Per‑asset strategy artifacts for isolated deployment.
- Periodic re‑tuning to keep drift under control.
Adaptive Code Factory (Vision)
We’re working toward a system that trains itself to produce and update strategy code aligned with current regimes — a disciplined “code factory.” How it works is intentionally not disclosed.
- Generate and update per‑asset strategy clones.
- Retune and redeploy on a cadence to manage regime drift.
- Maintain reproducible artifacts and clear provenance.
Risk Engine (Prototype)
Signals are monitored in unified risk units. We study favorable/adverse excursions and progress toward levels to suggest realistic take‑profit bands — without exposing internals.
- Excursion analytics and progress metrics.
- Simple expected‑value views for context.
- Designed for portfolio‑level roll‑ups.
System Architecture
- Separation of collect → compute → serve.
- Isolated workers, structured logging, timeouts.
- Per‑asset catalogs for downstream consumption.
Quantitative Approach
A long‑horizon momentum view is blended with an adaptive mean and envelopes. Zone transitions and retests help infer phase and price‑space levels. Specific formulas are intentionally withheld.
- Signals begin bias‑agnostic; final bias is determined by level vs. price.
- Edge cases are handled explicitly to avoid false cues.
- No repainting: backtests run prefix‑by‑prefix.
Research only: nothing here is investment advice.
Build Status & Roadmap
- Now: collectors, zone engines, tuning experiments, and risk prototypes.
- Next: harden walk‑forward evaluation and per‑asset artifacts; strengthen reporting.
- Later: automate the adaptive code factory; expand instrument coverage.