Reamer Research
A research engine for your mid-frequency strategies, built to institutional standards. Test an idea against years of OHLCV bars, diagnose it trade by trade, sweep it for robustness, emit a full JSON report, and connect it live through a relay. One engine, one stable C interface (ABI), byte-identical every run.
Strategies in Python or C++
The engine is a precompiled C library, and the kit ships two ways to call it, both in source under the kit licence. The C interface is the supported surface; the bindings are reference code you can read, change, and build on.
A working binding, no build step
Write an on_bar method, read prices as numpy arrays, return orders. Python 3.8+ and numpy; load your data with numpy, pandas, or csv. Templates for buy-and-hold, mean reversion, ATR brackets and multi-asset, plus a runnable quickstart.
A C++ wrapper
RAII result handles and a callable strategy adapter over the C interface, with buy-and-hold, ATR bracket, Donchian breakout and rejected-order examples. Anything else that calls C, from Rust to Go, can link the library directly.
From idea to live
A strategy travels the whole research path here, from the first run against history to the handoff to live execution, every stage in one engine behind one stable C interface.
Idea against history
Run a strategy against years of OHLCV bars, intraday to multi-day, with realistic execution costs. This is the backtest.
Trade by trade
Every fill, order, closed round trip and per-trade return is returned, so you can see exactly where a result came from, not just the summary.
Robustness across the space
Vary parameters and inputs across many runs to separate a robust result from one that survived on a single lucky configuration.
Full JSON out
A complete, schema-versioned JSON document per run: summary metrics, drawdown, streaks, and the full trade and order logs your own tools read.
Live through a relay
reamer_relay_* sends the same orders your strategy returns in a backtest to a running Reamer Server and reads the fills back as position state. Indicators that read a lookback window become running state. RESEARCH_TO_SERVER.md in the kit runs one order to a fill.
Byte-identical output across repeated runs
A fixed rng_seed produces byte-identical output across repeated runs, the stochastic slippage and spread path included. Determinism is a property of the seed, not of an absence of randomness.
Measured on AMD EPYC 9575F, single-threaded engine. Verify absolute numbers on your own hardware.
Integrated by Claude Code in about 13 minutes
Claude Code (Claude Sonnet) was given the Reamer Research kit and nothing else: no source code, no web access, no outside repos, no earlier exposure. Working from the C header, the execution specification, the result schema and the reference Python code, it built a multi-asset statistical-arbitrage strategy from scratch, showed that a second run gives byte-identical results, swept the strategy across signal and cost settings, traced the worst trade, and wrote the full JSON report, validated against the schema. A wall clock runs throughout: about 13 minutes of real time.
The kit documents are enough for a coding agent to do the wiring, so the learning curve is the agent's, not yours. You describe the strategy and the data; the agent writes the integration.
What happens in the recording
Chapter times are video time. The recording is sped up; the wall clock in the frame shows real elapsed time. Recorded 12 September 2026 against Reamer Research 4.2.1. Watch page → · YouTube →
- 0:00The task
- 0:20Verify, extract and read the kit
- 0:40Activate the licence, run the Python quickstart
- 1:00Choosing data for a pairs strategy
- 1:20Extending the Python binding
- 2:00Pair data and the strategy
- 2:20Run twice: byte-identical results
- 2:40Robustness sweep: signal and cost grids
- 3:20Forensics on the worst trade
- 4:00Full result JSON, validated against the schema
- 5:20What was built
- 5:40Recap: determinism, sweep, schema
- 6:00Deactivating the licence
A US-equity universe at full depth: 10,000 tickers × 20 years, 50M bars, 990,000 take-profit/stop-loss bracket orders and 990,000 trades, backtested in 29.1 s on AMD EPYC 9575F. Method in the whitepaper; the raw output is published.
One call across a stable C ABI, version 7. Feed bars and a strategy, receive a fully-resolved result set. Bars come from memory or from memory-mapped .bin files, so a dataset larger than RAM runs.
What's in scope
The engine models OHLCV bars for mid-frequency strategies, on data you bring. Platforms: linux_x86_64 (glibc 2.35+) and macOS arm64, ABI 7.
Per seat, per year. Start with a 30-day trial at $225; key and kit arrive by email within minutes.
Start a 30-day trial