HivemindOS manual

Quant Research Swarm

The quant research swarm is research-only: it has no order, broker, wallet, or money-movement path and cannot enable live trading. It tests a set of market ideas and keeps the evidence needed to reject weak ones.

Agents propose ideas and signals. A separate simulation engine runs cost-aware backtests, and an independent validator runs the statistics. A strategy passes only if every gate passes, and passing means only that it is ready for human review — not that it will make money.

What you provide

A reviewed, point-in-time dataset with its source and as-of dates.
How you handled survivorship bias and price adjustments.
One or more testable ideas, each with an economic reason.
Signal settings, with an execution delay of at least one bar.
Commission, slippage, and borrow cost assumptions.
Market returns and standard factor series (market, size, value, profitability, investment, momentum, and low volatility) for the independent checks.

Each request covers one symbol, so time order stays clear. You can test several strategies on the same history. Build portfolios from separately checked series.

What gets tested

Every run checks: enough observations, significance, bootstrap confidence, false discoveries, out-of-sample decay, overfitting, placebo signals, deflated Sharpe, alpha beyond known factors, market-regime behavior, and agreement between the engine and the validator.

Missing data fails the run. You can make thresholds stricter but never looser than the built-in minimums.

Run it

Check the installed policy:

hive-quant-research policy

Run a reviewed request file:

hive-quant-research run --request <request.json>

To run it on a schedule, set HIVEMINDOS_QUANT_RESEARCH_REQUEST to the reviewed request file, attach the hive-quant-research skill in Scheduler, pick a cadence, and approve its local file access. Scheduled runs stay local and research-only.

Results

Each run writes a manifest, a report, and separate candidate, backtest, and validation files. Hashes show whether a later result came from the same inputs. Runs are never overwritten, and failed runs keep their reason.