Three production systems running on a self-hosted Kubernetes cluster. The Crypto Pipeline section is wired live — data refreshes every fifteen minutes from a read-only Postgres replica. The other two are described in narrative, with live dashboards arriving as they leave development.
A distributed quantitative trading system. Real-time cryptocurrency data ingestion, technical analysis, signal generation, and automated paper trading — processing hundreds of symbols simultaneously on a bare-metal Kubernetes cluster.
Real-time USDT perpetuals streamed across multiple timeframes (1m to 1d). Tiered symbol selection by volume. Idempotent batch processing for fault tolerance.
Regime-aware signal engine. Market regime classification (trend, mean reversion, range) drives strategy selection at runtime — multiple strategies per regime, each calibrated to its risk profile. Volatility-scaled stop-loss and take-profit sizing.
Position limits, per-trade risk caps, and a daily loss circuit breaker. Automated execution with real-time P&L tracking and Telegram alerting across critical, warning, and info severities.
| Cointegration Pair | p-value | Half-Life · h | Correlation | Updated |
|---|
A two-layer macro intelligence pipeline. The sentiment engine collects macro-economic and social signals and composes them into a unified fear/greed index with regime classification. The indicator engine ingests a curated catalog of tier-1 macro series — rates, inflation, labor, growth, FX, credit, commodities, volatility, equities — and exposes both raw observations and a set of derived metrics (YoY/MoM changes, spreads, ratios, z-scores) for downstream analysis. Both write to the same TimescaleDB.
Two ingestion layers feed the same TimescaleDB. The sentiment layer pulls VIX, DXY, credit spreads and yield curves from Yahoo Finance and FRED, GDELT global news sentiment, Google Trends across thirteen economic keywords, YouTube trending across five countries, and Reddit (gated on API approval). The indicator layer maintains a curated catalog of seventy-three tier-1 macro-economic series across eighteen categories — monetary policy, Treasury yield curve, inflation (CPI / PCE / PPI), labor markets, growth and activity, credit conditions, commodities, FX, volatility, equities, ETFs, and housing — addressed through a normalised series_id namespace and refreshed at the cadence of each source.
The sentiment scorer composes a unified fear/greed index from seven weighted sources — VIX (25 %), credit spreads (15 %), news tone (15 %), policy uncertainty (15 %), yield curve (10 %), social sentiment (10 %), Google Trends (10 %) — and assigns a five-level regime classification from extreme fear to extreme greed. The indicator layer additionally computes derived metrics on top of the raw observations: year-on-year and month-on-month changes, cross-asset spreads (real 10Y yield, HY–IG OAS, term-structure premia), ratios (copper–gold for growth, NDX–RUT for rotation), and rolling z-scores and percentiles for context. All derivations recompute after each observation cycle.
Both pipelines write to TimescaleDB hypertables with per-chunk columnar compression past the active edit window and time-bucketed retention policies; the sentiment ingestors additionally use a Kafka dual-write pattern for redundancy. The schema migration from the v1 legacy layout to the v2 canonical model ran in parallel for a multi-week validation window before any legacy table was decommissioned — change-over without an outage event. Ansible-automated ingestor deployments, run-log records per cycle, and Telegram alerting on failure.
A large-scale data analytics engine that ingests billions of data points and runs quantitative analysis to uncover statistical relationships across crypto markets and macroeconomic signals.
Engle-Granger cointegration with half-life filtering, spectral analysis grounded in Random Matrix Theory, principal component decomposition, asset-graph network analysis with community detection, and regime classification. The statistical backbone behind every strategy and signal in the stack.
Return forecasting model in alpha phase — feature engineering and labeling pipelines currently producing tier-1 symbol predictions. Walk-forward validation with parameter stability testing gates promotion to production. Cross-pipeline signal fusion brings macro-sentiment regimes into the crypto strategy selector.
Planned next phase: PCA-derived factor models for systematic risk decomposition, Bayesian strategy ensembling weighted by recent regime fit, and PnL attribution that separates alpha from market beta and funding-rate drift. The framework that turns aggregate performance into actionable allocation signals.