This document captures proposed features, research directions, and improvements for AgentQuant. Ideas are organized by category and priority. React on GitHub issue #25 to vote on what matters most.
These must be solid before any real money is deployed.
- Description: Simulate 10k price paths and measure strategy robustness across scenarios
- Why: Backtest ≠ reality; reveals if strategy is robust or just lucky on historical data
- Impact: 🔴 CRITICAL — Can't ship live without this
- Effort: 2-3 days
- Acceptance: Generate confidence intervals on Sharpe; validate on synthetic market regimes
- Description: Implement optimal Kelly % with confidence intervals; dynamically adjust position size
- Why: Current agent has no bet sizing; Kelly is mathematically optimal for long-term growth
- Impact: 🔴 CRITICAL — Prevents overleverage and catastrophic losses
- Effort: 2 days
- Acceptance: Position size varies with strategy confidence; Sharpe-responsive
- Description: Trigger parameter re-learning when market regime shifts (VIX spike, regime score drop)
- Why: "Market regimes change; today's optimal parameters may not work tomorrow"
- Impact: 🔴 CRITICAL — Prevents strategy staleness
- Effort: 3 days
- Acceptance: Auto-trigger when regime score crosses threshold; regenerate parameters
- Description: Track if live Sharpe drifts >20% from backtest; auto-alert
- Why: Early warning system for strategy failure in production
- Impact: 🟠 HIGH — Operational alerting
- Effort: 1-2 days
- Acceptance: Dashboard shows live vs. backtest Sharpe; alerts on >20% drift
Better hypothesis generation and scoring.
- Description: Score hypotheses by "consensus" (multiple papers agree) vs. "contrarian" (unique insight); measure which wins
- Why: Contrarian ideas might have higher Sharpe but higher risk; consensus ideas more robust
- Impact: 🟠 MEDIUM — Refines hypothesis ranking
- Effort: 3-4 days
- Connects to: #24 (Hypothesis benchmarking)
- Acceptance: Hypotheses ranked by novelty score; report consensus vs. contrarian Sharpe
- Description: Generate hypotheses specific to seasonal patterns (Q1 effects, holiday trading, earnings season)
- Why: Many strategies exploit seasonal anomalies; worth formalizing
- Impact: 🟠 MEDIUM — Discovers new alpha sources
- Effort: 2-3 days
- Acceptance: LLM generates seasonal hypotheses; backtest on multi-year data
- Description: Track hypothesis age; decay its weight over time; measure how fast ideas stop working
- Why: Market evolves; ideas that worked 5 years ago may be arbitraged away
- Impact: 🟡 LOW — Research interest
- Effort: 2 days
- Acceptance: Measure half-life of hypothesis effectiveness
- Description: Take winning parameters, perturb (±5-10%), backtest; explore neighborhoods of good ideas
- Why: Fine-tune winners; discover adjacent strategies
- Impact: 🟡 LOW — Local optimization
- Effort: 2 days
- Acceptance: Generates 3-5 mutations per winning hypothesis; measures improvement
Evolve from single-strategy to production-grade portfolio.
- Description: Given 3-6 strategies with different regimes, find optimal allocation (min correlation, target Sharpe)
- Why: Portfolio > single strategy; diversification reduces volatility
- Impact: 🟠 MEDIUM — Scales to production
- Effort: 4-5 days
- Connects to: Next phase after single-strategy agent is solid
- Acceptance: Multi-strategy ensemble outperforms individual strategies
- Description: Detect current regime in real-time; automatically weight strategies (high-vol → volatility-aware; low-vol → mean reversion)
- Why: Each strategy thrives in different conditions; smart routing maximizes total Sharpe
- Impact: 🟠 MEDIUM — Improves adaptive returns
- Effort: 3 days
- Acceptance: Router weights strategies by regime; total Sharpe > any single strategy
- Description: Measure correlation of returns across strategies; identify diversifying pairs; test ensemble performance
- Why: Due diligence for portfolio construction
- Impact: 🟡 LOW — Infrastructure
- Effort: 1-2 days
- Acceptance: Correlation matrix shows diversification opportunities
- Description: Add puts/collar hedges during drawdown periods; optimize hedge ratio vs. cost
- Why: Hedge tail risk; reduces max drawdown
- Impact: 🟡 LOW — Risk management
- Effort: 3-4 days
- Acceptance: Hedged portfolio has lower max DD; Sharpe adjusted for hedge cost
Validate against public standards.
- Description: Run agent on same data as academic papers ("ML for trading" datasets); publish results vs. benchmarks
- Why: Credibility through comparison
- Impact: 🟠 MEDIUM — Academic credibility
- Effort: 2-3 days
- Acceptance: Published comparison table vs. paper baselines
- Description: Systematically disable each component (tools, ensemble, memory, etc.); measure impact on Sharpe; identify what matters
- Why: Scientific rigor; understand value of each component
- Impact: 🟡 LOW — Research transparency
- Effort: 3 days
- Acceptance: Report showing % Sharpe contribution per component
- Description: Test harness trained on SPY/QQQ against other markets (crypto, commodities, bonds); measure transfer performance
- Why: Broader applicability; is this market-specific or general?
- Impact: 🟡 LOW — Generalization research
- Effort: 2-3 days
- Acceptance: Strategies perform at >0.4 Sharpe in new markets
- Description: Return confidence intervals on Sharpe predictions, not point estimates ("0.58 ± 0.12" vs "0.58")
- Why: Honest reporting; stakeholders deserve uncertainty bounds
- Impact: 🟡 LOW — Better risk communication
- Effort: 2 days
- Acceptance: All metric outputs include confidence intervals
Make strategies understandable and debuggable.
- Description: LLM explains why a strategy works in current regime ("Mean reversion wins because Vol is elevated and momentum is exhausted")
- Why: Trust requires understanding; traders need to know why they should trade
- Impact: 🟠 MEDIUM — Interpretability & trust
- Effort: 2-3 days
- Acceptance: Each strategy includes generated English explanation
- Description: Convert learned parameters to interpretable rules ("If VIX > 75th AND RSI > 70, then shorter windows win")
- Why: Interpretability; regulatory/compliance requirements
- Impact: 🟡 LOW — Compliance & transparency
- Effort: 3 days
- Acceptance: Extract decision tree; validate against parameter choices
- Description: Generate video/animation of backtest with price chart + signals + P&L over time
- Why: Humans understand visuals better than JSON; debug faster
- Impact: 🟡 LOW — Developer experience
- Effort: 3-4 days (with video library)
- Acceptance: Generate .mp4 showing backtest with annotations
- Description: Track hypotheses that failed and why; learn patterns in failure ("Mean reversion fails when Vol is increasing, not just high")
- Why: Learn from failures; avoid repeating mistakes
- Impact: 🟡 LOW — Negative feedback learning
- Effort: 2 days
- Acceptance: Failure archive grows; negative patterns extracted
Moving from backtest → paper trading → live execution.
- Description: Run live with real-time data but don't execute; track if simulated fills match backtest assumptions
- Why: Bridge between backtest and live; validate assumptions hold in real markets
- Impact: 🔴 CRITICAL (when going live) — De-risks transition
- Effort: 4-5 days
- Acceptance: Paper trading Sharpe within 15% of backtest
- Description: Integration with Interactive Brokers, Alpaca, etc.; handle fills, slippage, latency
- Why: Execute in real markets
- Impact: 🔴 CRITICAL (when going live) — Operationalization
- Effort: 5-7 days
- Acceptance: Successfully execute trades and receive fills
- Description: Daily loss limits, max leverage, max concentration; enforce at execution
- Why: Prevent catastrophic losses; mandatory risk management
- Impact: 🔴 CRITICAL (when going live) — Risk containment
- Effort: 2-3 days
- Acceptance: Risk limits block orders that violate constraints
- Description: Slack/email for: strategy signals, regime changes, drawdowns, Sharpe drift
- Why: Monitor 24/7; catch problems early
- Impact: 🟠 MEDIUM (when going live) — Operational alerting
- Effort: 1-2 days
- Acceptance: Receive alerts for all material events
Better measurement of strategy quality beyond Sharpe.
- Description: Track multiple metrics (Calmar = Sharpe / max DD, Sortino = Sharpe on downside volatility, Omega = upside/downside ratio)
- Why: Sharpe can be gamed; other metrics measure different risk aspects
- Impact: 🟠 MEDIUM — Robust evaluation
- Effort: 2 days
- Acceptance: Dashboard shows 5+ metrics per strategy
- Description: Max DD, recovery time (avg time to recover after DD), DD frequency, DD distribution
- Why: Max drawdown is often the real constraint (psychological, risk management)
- Impact: 🟠 MEDIUM — Risk characterization
- Effort: 1-2 days
- Acceptance: Drawdown dashboard available; identify worst-case scenarios
- Description: 95% VaR, CVaR (Conditional VaR = expected loss beyond VaR); measure left-tail exposure
- Why: Regulatory requirement; quantifies tail risk
- Impact: 🟡 LOW — Risk measurement
- Effort: 2 days
- Acceptance: VaR metrics reported alongside Sharpe
- Description: Measure alpha vs. benchmark (SPY); "How much alpha is the strategy generating?"
- Why: Standard measure; compare passive returns to active strategy
- Impact: 🟡 LOW — Performance attribution
- Effort: 1 day
- Acceptance: IR reported; helps justify active management
Make the agent smarter and more autonomous.
- Description: Agent generates hypothesis, backtests, writes critique of why it worked/failed, stores learning
- Why: Agent learns from experience; improves hypotheses over iterations
- Impact: 🟠 MEDIUM — Autonomous learning
- Effort: 3-4 days
- Acceptance: Agent explicitly critiques own hypotheses; improves over time
- Description: Generate pairs of hypotheses (similar parameters, different Sharpe); learn what differentiates winners
- Why: Teach agent which parameter choices matter
- Impact: 🟡 LOW — Machine learning research
- Effort: 3-4 days
- Acceptance: Agent identifies key differentiators (e.g., "window length matters more than smoothing")
- Description: When proposing new idea, retrieve 3-5 similar historical hypotheses from memory; use as examples to Claude
- Why: In-context learning; Claude reasons better with examples
- Impact: 🟡 LOW — Prompt engineering
- Effort: 2 days
- Acceptance: Hypotheses include historical reference; Claude cites examples
- Description: LLM generates v1 hypothesis, we backtest, give feedback, LLM refines (2-3 iterations)
- Why: Iterative hypothesis improvement; find better ideas through refinement
- Impact: 🟡 LOW — Hypothesis quality
- Effort: 2-3 days
- Acceptance: Refined hypotheses outperform v1 baseline
See something exciting? Here's how to move it forward:
- React on GitHub issue #25 with 👍 to vote on what you want
- Comment with ideas — Add missing directions or variants
- Create an issue — When ready to build, reference this ROADMAP
- Link to related work — Connect to existing #19-24 issues
If you want maximum impact with limited time:
-
Monte Carlo Path Analysis (Credibility & Safety)
- Validates strategy robustness; can't skip before live trading
- 2-3 days; blocks 🚀 features
-
Portfolio Optimizer (Multi-Strategy & Portfolio)
- Natural scaling; single-strategy is research, multi-strategy is production
- 4-5 days; high ROI
-
Skepticism Scoring (Hypothesis Intelligence)
- Research-grade; differentiates signal from noise
- 3-4 days; improves hypothesis quality
Last updated: 2026-08-28
Status: Ideas stage (awaiting prioritization & issue creation)
Feedback: Open a comment or PR to update this roadmap