Architecture

wickra-backtest is a streaming-native, event-driven backtester built on the wickra-core indicator kernels. A strategy is data (a JSON spec), so the same strategy runs identically in ten languages and a backtest matches a live run by construction.

Crates

Crate	Role
wickra-backtest-core	The engine: spec DSL, indicator registry, rule evaluation, portfolio accounting, execution model, metrics, report and the streaming/batch runners.
wickra-backtest-data	Data loaders (CSV / JSON / JSON Lines / Parquet) and transforms (resampling, Renko / Kagi / Point-and-Figure bars).
wickra-backtest	A thin facade re-exporting the public API.
wickra-backtest-cli	The wkbt command-line backtester.
wickra-backtest-bench	Criterion throughput benchmarks.
bindings/*	Python, Node.js, WASM, C ABI, C#, Go, Java and R language bindings.

Data flow

candles ──▶ StreamingBacktest::step(candle, feeds)
                 │
                 ├─ 1. fill any pending order (look-ahead-free: on this bar's open)
                 ├─ 2. update each registered indicator from a BarInput
                 │      (candle + optional reference / derivatives / order-book /
                 │       trade / cross-section feeds)
                 ├─ 3. evaluate intrabar stop-loss / take-profit / trailing /
                 │      liquidation along the conservative O→H→L→C path
                 ├─ 3b. charge perpetual funding to an open position
                 ├─ 4. mark-to-market and push an equity point
                 └─ 5. evaluate entry / exit / short rules → queue an order
                        (filled at step 1 of a later bar, after latency)
            finish() ──▶ close any open position ──▶ BacktestReport

The historical run / run_with_capital is exactly this loop fed from a slice; pointing step at a live feed turns the same engine into the live bot. That single code path is why a backtest equals a live run.

Key design decisions

• Look-ahead bias is structural. A signal computed on a bar's close fills on the next bar's open by default; the optimistic same-bar close fill is explicit opt-in. The engine never reads a future bar.
• The registry is generated. tools/gen_registry.py parses the wickra-core indicator sources and emits registry.rs, so the 495 backtestable indicators stay in lock-step with the library. Each is wrapped behind a uniform, object-safe EvalIndicator the engine drives from a BarInput.
• Feeds are an extensible bundle. Feeds { reference, deriv, orderbook, trades, cross_section } threads microstructure context to the indicators that need it; absent feeds are simply None. run_json marshals one JSON request (candles + spec + feeds) so every binding exposes the full surface without marshalling variable-length feed arrays across the FFI.
• One kernel, ten languages. Every binding calls the same engine and returns the same JSON report. The golden corpus (golden/) pins this byte-for-byte for the OHLCV path and all four microstructure feed families.

Trust boundaries

The engine is pure computation: no network, no order placement, no API keys. Inputs (specs, candles, feeds) are untrusted JSON; parsing and evaluation never panic (see tests/properties.rs), always returning a Result. See THREAT_MODEL.md.