Four-block strategy files
Every strategy is a data, params, one-or-more signal, and portfolio file. The structure is the type system: the compiler knows exactly what each block must produce.
Features
Everything in the box.
sigc is a compiler, a runtime, a cache, a backtester, and a daemon — one Rust workspace, six crates. Here is the whole surface.
Compiler
The .sig language
Compile-time shape & calendar checks
Mismatched indices, wrong dtypes, and unresolved signal references fail at parse time — before a single row of data is read.
Composable signals
Reference signals by name and combine them with explicit arithmetic: 0.4 * momentum + 0.3 * value + 0.3 * quality. No hidden joins.
Language server (sig_lsp)
A dedicated LSP crate powers a VS Code extension with diagnostics, completion, and hover for .sig files.
Execution
The runtime
120+ vectorized operators
ret, lag, zscore, rank, winsor, rolling_mean, ema, rsi, macd, atr, vwap, neutralize and more — the surface area a factor researcher actually uses.
Polars / Arrow columnar engine
Signals lower to an operator graph evaluated lazily over Arrow columns, with Rayon parallelism.
SIMD kernels
AVX2 / AVX-512 kernels where the CPU supports them, for the hot numeric paths.
Custom signals
The operator set is not fixed: define your own named signal blocks and share factor libraries across teams.
Cache
Reproducibility
Content-addressed cache
IR and intermediate results are keyed by blake3 content hash and stored in a sled cache.
Byte-for-byte determinism
Identical .sig source, params, and inputs return identical outputs — down to the last basis point.
Microsecond warm hits
Rerunning an unchanged strategy hits the cache and returns in microseconds instead of recomputing.
Backtester
Backtesting
Portfolio constructors
rank(...).long_short(top, bottom, cap) for ranked books; scale_vol(signal, target_ann_vol, lookback) for vol-targeting.
Transaction-cost models
Compose costs: tc.bps(5) + slippage.model("square-root", coef=0.1) for commission plus square-root impact.
Rebalance cadence & benchmark
Explicit rebalance schedule, a date range, and an optional benchmark such as SPY for attribution.
Standard metrics
sigc run prints Total Return, Sharpe Ratio, Max Drawdown, and Turnover by default.
Deploy
Production
One binary, daemon mode
sigc daemon is a long-running process that owns the cache and serves clients over nng REQ/REP on tcp://127.0.0.1:7240.
Risk posture as config
sigc.yaml declares circuit breakers (max drawdown, max position, kill switch) and per-minute order rate limits.
Observability
Prometheus /metrics endpoint, structured JSON audit logs, and Slack alerts on anomalies.
Brokerless architecture
No Redis, no Kafka, no RabbitMQ — the daemon is the single cache owner; clients retry with exponential backoff.
Ecosystem
Integration
Rust embedding
Depend on the sigc crate, call Strategy::from_file, override params with with_param, and read weights/returns as Vec<f64>.
Python bindings (pysigc)
Stay in Jupyter but get the compiler’s guarantees on the underlying signal computation.
Data connectors
Load parquet, CSV, S3, GCS, or Postgres; Alpaca for paper and live trading per the README.
Six-crate workspace
sigc, sig_compiler, sig_runtime, sig_types, sig_cache, and sig_lsp — depend on exactly what you need.
See it end to end
The quickstart compiles and backtests a strategy in five minutes. The architecture page explains how the pieces fit.