Project · Agents in markets · July 2026
AlphaAgent: a multi-agent platform for markets
Markets are the harshest long-horizon environment we know: noisy feedback, regime shifts, and a scoreboard that settles every day whether you like it or not. That's exactly why we build there.
AlphaAgent is our multi-agent trading research platform, covering the full arc from alpha discovery to live execution. A team of specialized agents — data, prediction, alpha, sentiment, macro, risk, and execution — coordinate over an async message bus with explicit reasoning chains, so every decision the system makes can be traced from observation to conclusion. One side of the platform hunts for edge: machine-learned price models, academic alpha factors, and options strategies. The other side turns conviction into positions: a momentum portfolio executed through a brokerage API behind multiple safety gates, verified after every trade, and written up in a daily research report.
+53.8%
AlphaAgent live paper account, since March 2026
+13.5%
S&P 500 over the same period
4.0×
the index's return — every trade logged and verified
Paper trading on a live market feed. Research system — not investment advice.
How it works
From raw ticks to verified positions
One auditable loop: discover, decide, act, verify, report
01 — Ingest
Real-time and historical market data, options chains, and news — fetched in parallel across a 150+ name universe.
02 — Macro & regime
Market indices, 11 sectors, and 7 investment themes scored into a bull–bear regime signal, blended with an LLM macro synthesis.
03 — Predict
An ensemble of learned price models — LSTM, Transformer, and gradient-boosted trees — each reporting calibrated confidence.
04 — Score alpha
Academic 101-Alphas-style factors plus custom signals, evaluated by information coefficient and half-life; LLM news sentiment per ticker.
05 — Fuse
Regime- and confidence-weighted fusion into one conviction score, with every step recorded in an explicit reasoning chain.
06 — Act
Momentum portfolio construction and options strategies — executed through guarded brokerage orders with post-trade verification.
07 — Report
A daily research report with macro assessment, top picks, decision log, and live portfolio performance, delivered by email.
scroll →
The discovery side: hunting for edge
- Learned price models. LSTM, Transformer, and XGBoost forecasters run as an ensemble, with technical indicators and alpha factors as features and per-model confidence carried through to the fusion layer.
- Alpha factors, evaluated like a quant desk. Implementations from the academic "101 Alphas" catalog alongside custom factors — each judged by information coefficient, decay half-life, and cross-correlation before it earns weight in the score.
- Options as a first-class strategy. Black–Scholes pricing with full Greeks and an implied-volatility solver feed a strategy generator that proposes spreads, straddles, and iron condors matched to the volatility regime.
The execution side: built like infrastructure
- One source of truth for strategy math. The same portfolio-construction code drives backtests, daily reports, and live orders — the three can never silently drift apart.
- Hard risk rules, not suggestions. Per-position and per-sector caps, minimum cash reserves, notional limits, pending-order and buying-power guards, and post-trade verification on every execution.
- Honest backtesting. Next-day execution to kill look-ahead bias, transaction costs charged both ways, and per-year sweeps against both the S&P 500 and Nasdaq.
- Live, continuous evaluation. The execution loop has traded a live paper account since March 2026, up 53.8% against the S&P 500's 13.5% over the same period — every trade logged, performance docs generated automatically from the broker API.
Why agents, and why reasoning chains
The interesting problem isn't any single signal — it's coordination. AlphaAgent's agents communicate over an async message bus with pub/sub topics and request/response semantics, and every decision is recorded as an explicit observation → analysis → reasoning → conclusion chain. When the system takes a position, you can read exactly why. That auditability is the point: an agent acting autonomously with real consequences needs hard guardrails, verified execution, and a decision trail — the same properties every long-horizon agent will need. (AlphaAgent is a research system, not investment advice.)
Curious about the fusion architecture, the factor library, or the guardrail design? Write to research@r2m.ai.