Building crypto portfolios with agentic AI

Alessandro Piergallini; Antonino Castelli; Paolo Giudici

REVIEW 4 major objections 5 minor 13 references

Building crypto portfolios with agentic AI

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that a rolling 30-day Sharpe-maximizing portfolio, built by a collaborative multi-agent system, beats both equal-weight and one-shot mean-variance portfolios on the top ten cryptocurrencies from 2020 to 2025.

desk verdict A well-described agentic portfolio pipeline undermined by internally inconsistent reported metrics; the architecture is worth a look, the numbers are not yet trustworthy. read the letter →

arxiv 2507.20468 v1 pith:PRJYEHXU submitted 2025-07-11 q-fin.PM cs.LG

Antonino Castelli , Paolo Giudici , Alessandro Piergallini This is my paper · ORCID

classification q-fin.PMcs.LG

keywords portfoliooptimizationcryptocurrencymulti-agentsystemrollingwindowSharperatioModernTheoryagenticAImean-variance

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that an adaptive, agent-based pipeline can construct crypto portfolios that beat both naive equal weighting and one-shot mean-variance optimization. The central comparison is between a static optimizer and a 30-day rolling-window optimizer that rebalances the top ten cryptocurrencies by market capitalization. On daily data from August 2020 to March 2025, the rolling strategy reports higher risk-adjusted returns in-sample (Sharpe 1.00 versus 0.60 for static optimization) and out-of-sample (0.72 versus 0.36), with lower annualized volatility. If those numbers are reliable, the practical lesson is that frequent rebalancing helps in volatile markets where correlations and volatilities shift.

What carries the argument

The load-bearing mechanism is a 30-day rolling-window Sharpe-ratio maximizer with no short selling. Every 30 days, the optimizer takes the most recent 30 days of daily returns, solves for the fully-invested non-negative weights that maximize the annualized expected-return-minus-risk-free over volatility, applies those weights for the next 30 days, and then repeats. Around this optimizer the paper builds a multi-agent pipeline whose dedicated agents handle data loading, cleaning, train/test splitting, metric computation, benchmarking against the S&P 500, and a file-integrity check, with a final reporting agent that interprets the tool outputs. What makes the argument go through is the contrast between this recalculating optimizer and a static mean-variance optimizer applied once on the training set.

What would settle it

Recompute all in-sample and out-of-sample metrics directly from the data and the published weight rules, then subtract realistic per-trade transaction costs (spreads and fees) from the rolling strategy. If the rolling strategy's out-of-sample Sharpe no longer exceeds the static strategy's, or if an independent implementation yields values close to 0.36 rather than 0.72, the paper's central outperformance claim collapses.

Watch

Extended reading notes

Core claim

The core claim is that portfolio adaptivity drives outperformance in crypto markets. The paper's dynamic strategy, called Crew B, recomputes Sharpe-maximizing weights every 30 days, using the trailing 30-day window, under no-short-selling and full-investment constraints. Over the ten largest cryptocurrencies from 2020-08-20 to 2025-03-13, this strategy is reported to achieve an in-sample Sharpe ratio of 1.00 and an out-of-sample Sharpe ratio of 0.72, compared with 0.60 and 0.36 for the static mean-variance portfolio (Crew A) and 0.48 for the equal-weighted benchmark in-sample. The same rolling strategy also shows lower volatility (10% annualized in-sample and 11% out-of-sample) and keeps maximum drawdown near -15%, which the paper interprets as evidence that time-varying allocation is better suited to regime-shifting crypto markets.

Load-bearing premise

The central claim depends on the reported Sharpe ratios having been computed by the portfolio tools and faithfully reported, so that the 0.72-versus-0.36 out-of-sample gap is real rather than an artifact of the language models' summaries.

Editorial extensions

If this is right

For a monthly-rebalancing investor in the top-10 crypto universe, the paper predicts higher risk-adjusted returns from rolling 30-day Sharpe maximization than from a buy-and-hold mean-variance portfolio.
Static mean-variance portfolios become stale as crypto correlations and volatilities change; the paper's results imply that any fixed-weight crypto allocation should be treated as a benchmark, not a final strategy.
The modular agent pipeline is claimed to be scalable: adding assets, benchmarks, or cost models requires spawning new agents rather than rewriting the system.
Because transaction costs and tail-risk metrics such as CVaR are not modeled, the paper's own caveat is that part of the edge could be eroded in live trading by spreads, fees, and extreme-event losses.
If the tool-computed metrics are reproducible, the same architecture should be portable to other volatile asset classes or to crypto portfolios with more than ten tokens.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: the paper presents two sets of in-sample Sharpe values for the static optimizer (0.83 in the agent report, 0.60 in the main text); an outside reader should treat the precise magnitudes as provisional until the raw tool outputs are published.
Editorial inference: the same 30-day rolling machinery could be stress-tested on sub-periods such as the 2022 bear market versus the 2023-24 rally to see whether the outperformance is concentrated in specific regimes.
Editorial inference: adding a transaction-cost layer and testing rebalancing windows of 7, 14, and 60 days would turn the central claim into an actionable calibration rather than a single-window proof of concept.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Here's my take on arXiv:2507.20468. The genuinely new thing is the application: a Crew AI multi-agent pipeline that automates data loading, cleaning, portfolio optimization, and reporting for a top-10 crypto universe, with a clean comparison between equal-weight, static mean-variance, and 30-day rolling Sharpe maximization. The agent architecture is described in enough detail to be intelligible, and the authors are candid about limitations (transaction costs, heuristic window, missing tail risk). That part is respectable.

The problem is that the empirical core doesn't hold together. Section 5.1 reports equal-weight Sharpe 0.48 and Crew A Sharpe 0.60 in-sample. Table 3 and Report A give the same strategies Sharpe 0.53 and 0.83, with volatility 15% vs 12.4%. Out-of-sample, Section 5.2 says Crew A drops to 0.36, but Report C reports 0.69. These aren't rounding artifacts; they're 30-90% differences in the key numbers. Because no code or data is released, I can't tell which set of numbers came from the portfolio metrics tool and which came from the Final Report agent's narrative. That's a load-bearing problem: the paper's central claim that Crew B beats the alternatives rests on numbers that contradict each other.

I'm also unimpressed by the claim of 'significantly better performance' with no statistical test. With 334 out-of-sample daily observations, a simple block bootstrap or a Diebold-Mariano-style test would be trivial to run. The absence of any significance evidence makes the word 'significantly' unsupported.

What's not a problem: the underlying finance idea is standard rolling-window optimization, and the paper doesn't oversell it as new theory. The novelty claim is the agentic wrapper, which is legitimate as an application. The self-citations don't bother me; citing your own explainable-crypto work is fine.

Who would benefit from reading this? Someone working on LLM agents for financial workflows, especially in crypto, will find the architecture a useful template. But as a reported result, the numbers need to be recomputed from the stated rules before I'd trust any conclusion. The good news is this is fixable: release the code, reconcile the tables, and add a significance test. If the corrected numbers reproduce Crew B's 0.72 out-of-sample Sharpe, this becomes a credible applied paper.

My recommendation: send it to peer review, not because it's ready, but because the architecture and question are worth referee time, and the empirical issues are exactly what referees should catch. I'd require the code/data and consistency checks before acceptance.

Referee Report

4 major / 5 minor

Summary. The paper proposes a multi-agent system built with Crew AI to automate the construction and evaluation of cryptocurrency portfolios. It compares a static equal-weighted portfolio and a static mean-variance optimized portfolio (Crew A) against a 30-day rolling-window Sharpe-maximizing portfolio (Crew B), using daily data on ten large cryptocurrencies from August 2020 to March 2025. The authors report that Crew B achieves higher in-sample and out-of-sample Sharpe ratios than the static strategies, and they argue that this demonstrates the value of adaptive techniques and of modular, auditable multi-agent systems in financial automation. The manuscript includes a description of the data, tools, agents, in-sample and out-of-sample results, a comparative discussion, and limitations.

Significance. If the reported results were reliable, the paper would provide a practical demonstration of LLM-based multi-agent systems in crypto portfolio management, with an architecture that separates data handling, optimization, and reporting. The topic is timely and the modular design is described clearly. However, the contribution is essentially empirical: a single backtest on one historical period. The paper's central outperformance claim is undermined by internal contradictions in the reported metrics, the lack of any statistical uncertainty assessment, and the absence of released code or data. As a methodological contribution, the paper does not go beyond standard rolling-window mean-variance optimization; its value depends entirely on the credibility of the computed numbers, which the manuscript does not establish.

major comments (4)

[Section 5.1, Table 3, REPORT CREW A] The reported performance metrics for the same strategies disagree across the in-sample results, Table 3, and the agent reports. Section 5.1 reports Crew A in-sample Sharpe 0.60 and equal-weight Sharpe 0.48, while Table 3 lists Crew A Sharpe 0.83 and equal-weight Sharpe 0.53, with expected return and volatility also differing (7.5%/12.4% vs 10%/12% and 8%/15%). Additionally, REPORT CREW A gives the test-set Sharpe for the optimized static strategy as 0.69, whereas Section 5.2 reports Crew A out-of-sample Sharpe 0.36. These are not rounding differences; the same strategy is assigned Sharpe ratios that differ by roughly 30% to 90%. Because the abstract and the conclusion rest on the claim that Crew B outperforms Crew A, and because the reported baseline numbers are internally inconsistent, the central empirical claim is not supported as presented.
[Section 2, Section 5.3] The manuscript treats agent-generated reports as evidence of performance, but the Final Report Agent is explicitly described as using no tool ('its role is analytical and interpretative') and the reports in Section 5.3 are narrative outputs of the LLM agents. The paper never shows the raw output of the portfolio metrics tool for the strategies in question, so the reader cannot determine whether the headline in-sample and out-of-sample Sharpe ratios (e.g., Crew B's 1.00 and 0.72) are computed by the tools or produced by the agents' text generation. This design creates a risk that the reported figures are hallucinated or narratively embellished, and it is a load-bearing gap for a paper whose only substantive result is a set of performance numbers.
[Abstract; Section 5.3] The abstract and Section 5.3 state that Crew B 'significantly' outperforms the static strategies, but no statistical hypothesis tests, confidence intervals, standard errors, or cross-validation procedures are reported. The evaluation is based on a single 80/20 train/test split of one historical sample. In a highly volatile asset class such as cryptocurrencies, a single backtest cannot support the word 'significantly' without some measure of uncertainty. This is a load-bearing omission for the paper's central claim of superior risk-adjusted returns.
[Section 2; Section 6] The manuscript claims that the MAS design guarantees 'replicability' and is 'fully auditable by supervisors,' yet neither the code, the dataset, nor the agent prompts are provided. The file checker tool is described, but no outputs are shown. Without code or data, the empirical results cannot be independently verified, and the internal inconsistencies described in the first comment cannot be resolved by the reader. The limitations section mentions transaction costs, window length, and universe size, but it does not quantify these effects, and the paper does not provide a robustness check such as alternative window lengths or a different train/test split.

minor comments (5)

[REPORT CREW A, item 1] The text contains a typo: 'T rain Set' instead of 'Train Set'.
[Table 1] The row for SHIB shows a mean, median, and standard deviation of 0 due to rounding; the text notes this, but the table would benefit from scientific notation or a footnote stating the actual small values.
[Section 2, Figure 1] Figure 1 is referenced in the text but not included; the agent architecture is only described in prose, which makes the workflow harder to follow.
[References [1] and [5]] The introduction and conclusion rely heavily on the authors' own prior work ([1] and [5]) to frame the contribution; the manuscript should state explicitly what is new relative to those papers, rather than only citing them for claims about explainability and adaptivity.
[Section 2, Section 4.2] The choice of a 30-day rolling window and an 80/20 split are justified only heuristically; a sentence acknowledging the sensitivity of the results to these choices and pointing to a robustness check would strengthen the presentation.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor in-sample tautology; the central out-of-sample claim is an independent backtest, so no significant circularity.

fitted input called prediction [Section 4.1 (Crew A: Static Portfolio Analysis) and Section 5.1 (In-Sample Performance)]
"it applies a mean-variance optimization to determine the portfolio weights that maximize the Sharpe ratio... We observe that Crew A improves the Sharpe ratio from 0.48 (equal-weighted approach) to 0.60 (optimized weights approach)."

The weights reported as 'Crew A' are the solution of an in-sample Sharpe maximization, and the equal-weight portfolio is a feasible candidate in that optimization (weights sum to 1, each between 0 and 1). Therefore the optimized portfolio's in-sample Sharpe is guaranteed to be at least as large as the equal-weight portfolio's in-sample Sharpe by construction. The claimed in-sample improvement from 0.48 to 0.60 is thus a tautological consequence of the fitting procedure rather than an independent empirical finding. The out-of-sample comparison (0.36 vs 0.72) is not forced by this construction, so the paper's central adaptive-strategy claim retains independent content.

full rationale

The paper's central claim is an empirical backtest: Crew B's rolling 30-day Sharpe-maximizing allocation is compared with equal weight and static mean-variance portfolios on a train/test split. This comparison does not reduce to a restatement of the inputs. The one definitional element is in Section 5.1, where Crew A's in-sample Sharpe improvement over equal weight is guaranteed by the optimizer's objective of maximizing in-sample Sharpe with equal weight as a feasible candidate; reporting that improvement as a result is a fitted-input artifact. The out-of-sample Sharpe comparison is not forced and gives the paper independent content. The self-citations [1] and [5] support framing and future work but are not load-bearing: the outperformance conclusion rests on the reported backtest numbers, not on those citations. The Final Report Agent's self-assessment and the numerical inconsistencies between Section 5.1, Table 3, and the agent reports are correctness and reproducibility risks, not circularity under the definition used here. Score 2 reflects one minor tautological in-sample step with no load-bearing self-citation chain.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The analysis rests on standard portfolio-theory assumptions plus several hand-chosen settings (30-day window, 80/20 split, top-10 universe). The most fragile addition is the reliance on LLM agents to compute and narrate the metrics, which the internal inconsistencies call into question.

free parameters (4)

Rolling window length = 30 days
The 30-day lookback period for re-estimating Sharpe-optimal weights is chosen heuristically; the authors acknowledge alternative windows may change results (Section 6).
Train/test split = 80/20
The dataset is split into 80% training and 20% test on a chronological basis; the split is a modeling choice.
Asset universe = Top 10 by market cap
The ten assets are selected by market capitalization over the sample period; survivorship bias may be introduced.
Risk-free rate = Implicitly 0
The Sharpe ratio formula in Section 4.1 includes Rf but no value is specified, implying Rf = 0.

assumptions (5)

standard math Mean-variance optimization and Sharpe ratio are valid criteria for portfolio selection
Used throughout Section 4.
standard math Daily returns can be annualized by sqrt(252) scaling
Section 4.1 volatility formula.
domain assumption The 30-day history is sufficient to estimate expected returns and covariances
Crew B relies on 30-day windows (Section 4.2).
ad hoc to paper The top-10 cryptocurrency selection is representative and free of look-ahead bias
Section 2 defines the universe without specifying when market cap was measured.
ad hoc to paper LLM agents in Crew AI produce faithful calculations rather than narrative outputs
The Final Report Agents generate the performance tables in Section 5.3.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Building crypto portfolios with agentic AI." pith.science (2026). https://pith.science/paper/PRJYEHXU

@misc{pith2026250720468,
  author       = {Pith},
  title        = {Pith review of: Building crypto portfolios with agentic AI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PRJYEHXU}},
  note         = {Machine review of arXiv:2507.20468}
}

read the original abstract

The rapid growth of crypto markets has opened new opportunities for investors, but at the same time exposed them to high volatility. To address the challenge of managing dynamic portfolios in such an environment, this paper presents a practical application of a multi-agent system designed to autonomously construct and evaluate crypto-asset allocations. Using data on daily frequencies of the ten most capitalized cryptocurrencies from 2020 to 2025, we compare two automated investment strategies. These are a static equal weighting strategy and a rolling-window optimization strategy, both implemented to maximize the evaluation metrics of the Modern Portfolio Theory (MPT), such as Expected Return, Sharpe and Sortino ratios, while minimizing volatility. Each step of the process is handled by dedicated agents, integrated through a collaborative architecture in Crew AI. The results show that the dynamic optimization strategy achieves significantly better performance in terms of risk-adjusted returns, both in-sample and out-of-sample. This highlights the benefits of adaptive techniques in portfolio management, particularly in volatile markets such as cryptocurrency markets. The following methodology proposed also demonstrates how multi-agent systems can provide scalable, auditable, and flexible solutions in financial automation.

Figures

Figures reproduced from arXiv: 2507.20468 by the authors.

**Figure 2.** Bitcoin Prices and Returns Comparison (2020–2025) [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 9 canonical work pages

[1]

SSRN Electronic Journal (2021)

Babaei, G., Giudici, P.: Explainable artificial intelligence for crypto asset alloca- tion. SSRN Electronic Journal (2021)

work page 2021
[2]

Advances in Economics Management and Political Sciences 90, 91–97 (2024)

Chen, P.: Portfolio optimization based on rolling window using 5 stocks. Advances in Economics Management and Political Sciences 90, 91–97 (2024)

work page 2024
[3]

arXiv preprint arXiv:2409.17266 (2024)

Cheng, J., Chin, P.: LLM agent-based asset pricing models. arXiv preprint arXiv:2409.17266 (2024)

arXiv 2024
[4]

Algorithms 17(12), 570 (2024)

Espiga-Fern´ andez, F., Garc ´ ıa-S´ anchez,´A., Ordieres-Mer´ e, J.: A systematic ap- proach to portfolio optimization: A comparative study of reinforcement learning agents, market signals, and investment horizons. Algorithms 17(12), 570 (2024)

work page 2024
[5]

Physica A: Statistical Mechanics and its Applications 655, 130176 (2024)

Giudici, P., Piergallini, A., Recchioni, M.C., Raffinetti, E.: Explainable artificial intelligence methods for financial time series. Physica A: Statistical Mechanics and its Applications 655, 130176 (2024)

work page 2024
[6]

rep., National Bureau of Economic Research (2023)

Horton, J.J.: Large language models as simulated economic agents: What can we learn from Homo Silicus? Tech. rep., National Bureau of Economic Research (2023)

work page 2023
[7]

MAPS: Multi-agent Reinforcement Learning-based Portfolio Management System

Lee, J., Kim, R., Yi, S.W., Kang, J.: MAPS: Multi-agent reinforcement learning- based portfolio management system. arXiv preprint arXiv:2007.05402 (2020)

work page Pith review arXiv 2020
[8]

arXiv preprint arXiv:2404.11584 (2024)

Masterman, T., Besen, S., Sawtell, M., Chao, A.: The landscape of emerging AI agent architectures for reasoning, planning, and tool calling. arXiv preprint arXiv:2404.11584 (2024)

arXiv 2024

Show all 13 references

[9]

arXiv preprint arXiv:2103.00949 (2021)

Misheva, B.H., Osterrieder, J., Hirsa, A., Kulkarni, O., Lin, S.F.: Explainable AI in credit risk management. arXiv preprint arXiv:2103.00949 (2021)

2021 arXiv
[10]

arXiv preprint arXiv:2403.19735 (2024)

Park, T.: Enhancing anomaly detection in financial markets with an LLM-based multi-agent framework. arXiv preprint arXiv:2403.19735 (2024)

2024 arXiv
[11]

arXiv preprint arXiv:2409.20536 (2024)

Valdrighi, G., Ribeiro, A.M., Pereira, J.S.B., Guardieiro, V., Hendricks, A., Mi- randa Filho, D., Nieto Garcia, J.D., Bocca, F.F., Veronese, T.B., Wanner, L., Raimundo, M.M.: Best practices for responsible machine learning in credit scor- ing. arXiv preprint arXiv:2409.20536 (2024)

2024 arXiv
[12]

AAAI Conference on Artificial Intelligence (2025)

Xiao, Y., Sun, E., Luo, D., Wang, W.: TradingAgents: Multi-agents LLM finan- cial trading framework. AAAI Conference on Artificial Intelligence (2025)

2025
[13]

arXiv preprint arXiv:2407.18957 (2024)

Zhang, C., Liu, X., et al.: When AI meets finance (StockAgent): Large language model-based stock trading in simulated real-world environments. arXiv preprint arXiv:2407.18957 (2024). 12

2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

SSRN Electronic Journal (2021)

Babaei, G., Giudici, P.: Explainable artificial intelligence for crypto asset alloca- tion. SSRN Electronic Journal (2021)

work page 2021

[2] [2]

Advances in Economics Management and Political Sciences 90, 91–97 (2024)

Chen, P.: Portfolio optimization based on rolling window using 5 stocks. Advances in Economics Management and Political Sciences 90, 91–97 (2024)

work page 2024

[3] [3]

arXiv preprint arXiv:2409.17266 (2024)

Cheng, J., Chin, P.: LLM agent-based asset pricing models. arXiv preprint arXiv:2409.17266 (2024)

arXiv 2024

[4] [4]

Algorithms 17(12), 570 (2024)

Espiga-Fern´ andez, F., Garc ´ ıa-S´ anchez,´A., Ordieres-Mer´ e, J.: A systematic ap- proach to portfolio optimization: A comparative study of reinforcement learning agents, market signals, and investment horizons. Algorithms 17(12), 570 (2024)

work page 2024

[5] [5]

Physica A: Statistical Mechanics and its Applications 655, 130176 (2024)

Giudici, P., Piergallini, A., Recchioni, M.C., Raffinetti, E.: Explainable artificial intelligence methods for financial time series. Physica A: Statistical Mechanics and its Applications 655, 130176 (2024)

work page 2024

[6] [6]

rep., National Bureau of Economic Research (2023)

Horton, J.J.: Large language models as simulated economic agents: What can we learn from Homo Silicus? Tech. rep., National Bureau of Economic Research (2023)

work page 2023

[7] [7]

MAPS: Multi-agent Reinforcement Learning-based Portfolio Management System

Lee, J., Kim, R., Yi, S.W., Kang, J.: MAPS: Multi-agent reinforcement learning- based portfolio management system. arXiv preprint arXiv:2007.05402 (2020)

work page Pith review arXiv 2020

[8] [8]

arXiv preprint arXiv:2404.11584 (2024)

Masterman, T., Besen, S., Sawtell, M., Chao, A.: The landscape of emerging AI agent architectures for reasoning, planning, and tool calling. arXiv preprint arXiv:2404.11584 (2024)

arXiv 2024