In the volatile world of crypto, the only constant is change. This makes portfolio optimization a challenge — especially for traditional strategies that assume stability over time. A new study by Castelli, Giudici, and Piergallini offers a compelling solution: build your investment pipeline out of agents.
Using a modular Multi-Agent System (MAS) framework implemented in Crew AI, the authors compare two crypto portfolio strategies over the 2020–2025 period: one static and one adaptive. The system orchestrates specialized agents to ingest, clean, analyze, optimize, and report on daily crypto prices — all in a transparent and auditable way.
🧰 The Setup: Agents in the Loop
The MAS divides its pipeline into modular agents, each responsible for a single task:
- Data Loader: Fetches daily close prices of the top 10 cryptocurrencies
- Cleaner: Handles anomalies and missing values
- Splitter: Splits dataset into 80% train, 20% test
- Optimizer: Generates portfolio weights (static or rolling)
- Evaluator: Computes key metrics: Expected Return, Sharpe, Sortino, Volatility
- Report Writer: Produces human-readable summaries
Two separate crews are defined:
| Crew | Strategy | Optimization Logic | Rebalancing |
|---|---|---|---|
| A | Static | Mean-Variance | None |
| B | Dynamic | Rolling Sharpe-Max | Every 30 days |
📊 The Results: Adaptivity Wins
| Metric | Equal Weights | Crew A (Static) | Crew B (Rolling) |
|---|---|---|---|
| Expected Return | 8% | 10% | 10% |
| Volatility | 15% | 12% | 10% |
| Sharpe Ratio | 0.53 | 0.83 | 1.00 |
| Sortino Ratio | 0.75 | 1.10 | 1.30 |
| Max Drawdown | -20% | -15% | -15% |
Crew B’s rolling strategy doesn’t just look better on paper — it holds up out-of-sample as well. While Crew A’s Sharpe drops to 0.36 when exposed to new data, Crew B maintains a strong 0.72. This speaks to the robustness and generalization ability of the rolling approach.
🧩 Why Agentic AI Matters
The paper isn’t just about better returns — it’s about building a better system:
- Modular: Each agent (e.g., optimizer, cleaner) can be upgraded independently.
- Auditable: Agents log their decisions, making it easy for compliance and monitoring.
- Scalable: Want to add stablecoins, risk models, or benchmarks? Just spawn new agents.
These features make the MAS approach ideal not only for crypto, but for any market where transparency, speed, and flexibility matter.
🧪 Tradeoffs & Limitations
- Transaction Costs: Ignored in this study. Rolling rebalancing may incur fees.
- Top-10 Bias: Only considers the most capitalized coins.
- Window Size: 30-day lookback is a heuristic — not optimized.
- Tail Risk: No modeling of extreme events (e.g., CVaR).
Still, these are limitations of scope, not structure. The MAS framework could easily incorporate cost models, CVaR modules, or longer lookbacks via additional agents.
🔮 The Takeaway
The Crew AI system demonstrates that agentic infrastructure + adaptive strategy = superior crypto portfolios. It’s not just about beating equal weighting or static optimization. It’s about building an investment engine that adapts — modularly, transparently, and continuously — to a market where yesterday’s alpha is today’s beta.
For financial institutions eyeing crypto, this architecture offers a rare combination: performance uplift without black-box opacity.
Cognaptus: Automate the Present, Incubate the Future.