DATABASES Signal 224
Walsh: Multi-agent research pipeline with risk manager that can veto trades
For engineers building automated decision systems, this provides a concrete architecture pattern for combining multiple independent analysis agents with a risk gate that can override consensus. The backtest results demonstrate the inherent tradeoff: conservative risk thresholds that correctly identify high-uncertainty periods also cause missed opportunities in trending markets, resulting in underperformance versus buy-and-hold.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
The Risk Manager applies hard veto rules (confidence below 45% or agent signal spread ≥ 4) and soft downgrade rules (confidence 45–65% downgrades strong signals), which correctly avoided losses during volatile periods but caused the pipeline to exit all positions and hold cash for the entire month of October 2023, missing a subsequent NVDA recovery rally.
The current implementation uses rule-based agent stubs rather than LLM-backed agents, with per-agent accuracy calibration, cost/latency metrics, and meaningful confidence curves all pending future population once LLM agents replace the stubs.
The Portfolio Manager applies a consensus penalty that scales confidence down based on the spread between agent signals, meaning extreme disagreement (such as STRONG_BUY versus STRONG_SELL) halves confidence, and adversarial testing flagged 5 out of 18 test cases for overconfidence.
THE CLUSTER