MountainLion: Multi-Agent Crypto Trading LLM

• MountainLion is a multi-modal, multi-agent LLM system that integrates heterogeneous inputs—textual news, candlestick charts, and on-chain flows—to generate interpretable cryptocurrency trading insights.
• It orchestrates specialized agents for technical analysis, market dynamics, and semantic refinement, ensuring real-time report generation and responsive forecasting.
• A central reflection module adjusts prompt strategies and fusion weights based on historical performance, enhancing both transparency and adaptability in trading decisions.

Searching arXiv for the specified paper and a closely related context paper to ground the article. MountainLion is a multi-modal, multi-agent LLM system for interpretable and adaptive cryptocurrency trading. It is designed for a setting in which trading decisions depend on heterogeneous inputs, including textual news, candlestick charts, trading signal charts, and on-chain capital flows, and it addresses limitations attributed to prior deep learning and reinforcement learning approaches, notably heavy data requirements, reduced transparency from latent numerical encodings, weak adaptability under regime shifts, and limited investor-facing interpretability (Wu et al., 13 Jul 2025). The system coordinates specialized agents for technical analysis, market dynamics analysis, recommendation synthesis, and semantic refinement, while a central reflection module continuously evaluates historical signals and outcomes to refine prompt strategies, fusion weights, and evidence selection. This yields a framework in which financial reports, forecasts, and recommendations are generated, revised, and interactively adjusted in near real time.

1. Problem setting and design rationale

Cryptocurrency trading is characterized in the source paper as fast-moving, sentiment-driven, and heterogeneous across modalities. Decision quality depends on situational awareness across textual narratives, visual market signals such as candlestick “K-line” charts and trading signal charts, and on-chain capital flows. The motivating claim is that effective trading requires the reconciliation of technical price triggers with macroeconomic and regulatory context, rather than reliance on purely technical indicators or static model outputs (Wu et al., 13 Jul 2025).

Within that framing, MountainLion is positioned against earlier DL and RL approaches that require large labeled datasets and encode heterogeneous inputs into latent numerical vectors, often at the cost of transparency and adaptability. Those methods are further described as often static at inference time, with weak responsiveness to evolving sentiment and regulation. MountainLion instead adopts an LLM-based multi-agent decomposition, coupled with Retrieval-Augmented Generation and graph reasoning, so that diverse and up-to-date information sources can be processed in natural language and iteratively combined into investor-facing reports.

A central premise of the system is that interpretability is not an auxiliary property but part of the operational design. Reports are intended to explain why a signal or forecast is actionable, and the architecture is structured so that technical, contextual, and narrative elements remain separable before final synthesis. This suggests that MountainLion treats explanation as a first-class systems requirement rather than as a post hoc wrapper around a forecasting model.

2. Architecture and agent organization

MountainLion uses a layered architecture comprising a User Interface layer, a Core Business Logic layer, an AI Engine, and a Database Layer (Wu et al., 13 Jul 2025). The UI is described as a multilingual dashboard for alerts, chat, and visual analytics. Core Business Logic orchestrates multi-agent report generation, price forecasting, and news-driven recommendations. The AI Engine contains LLM agents, continuous real-time retrieval from exchanges and news sources, GraphRAG for semantic enrichment, lightweight ML predictors, and an LLM-powered coordination module. The Database Layer uses MySQL-backed storage for OHLC market data at multiple granularities, news, generated reports, and dense embeddings.

The agent design is specialized. The A1 Technical Analysis Agent computes classical indicators, including RSI, MACD, and Bollinger Bands, together with support and resistance zones and trend strength from OHLCV. The A2 Market Dynamics Agent synthesizes news sentiment, regulation signals, funding and capital flows, and social signals such as KOL activity. The A3 Trading Recommendation Agent integrates the outputs of A1 and A2 to produce short-, medium-, and long-term strategies. The A4 Semantic Agent performs LLM-driven coherence and narrative refinement to ensure lexical consistency and investor-facing clarity (Wu et al., 13 Jul 2025).

Inter-agent coordination is mediated by retrieval and reflective control. RAG and GraphRAG provide evidence and entity-linked context, while a Perplexity-based retriever augments draft reports with fresh external signals. A central reflection module evaluates rolling performance, including forecast accuracy and directional “win rate,” and uses those outcomes to refine fusion weights, prompt schemas, and retrieval scoring. The significance of this arrangement lies in its separation of evidentiary acquisition, analytic specialization, integration, and self-adjustment into explicitly named system components.

3. Modalities, representations, and feature extraction

MountainLion processes textual news, OHLCV-derived chart structure, and trading signal charts, but it does not specify a dedicated pixel-level vision backbone. Instead, the paper emphasizes indicator-based chart interpretation over raw-image computer vision (Wu et al., 13 Jul 2025). This is a technically important distinction: the system is multi-modal in the sense of integrating heterogeneous information channels, but its visual pathway is primarily structured through financial time-series representations rather than end-to-end learned visual encoders.

For textual news, the system performs continuous ingestion via APIs, RSS, and curated portals. Semantic annotation includes sentiment classification using LLM heads and named entity recognition over the entity types $$\{ \text{ORG}, \text{PER}, \text{EVT}, \text{CRYPTO} \}$$. These annotations feed a dynamic knowledge graph in which nodes represent documents and entities and edges encode associations via co-occurrence and context. The graph supports multi-hop retrieval and grounded summarization. The paper also describes preference-aware retrieval that infers user asset class, risk tolerance, and horizon, and then uses prompts for evidence selection (Wu et al., 13 Jul 2025).

For market data, OHLCV “K-line” data at 1d, 1h, and 5m granularities are stored in MySQL, and technical indicators are computed to capture trend, momentum, and volatility. The paper gives explicit definitions for the principal indicators:

$EMA_t = \alpha P_t + (1 - \alpha) EMA_{t-1}$

$$RSI = 100 - \frac{100}{1 + RS}, \quad \text{where } RS = \frac{\text{Average Gain}}{\text{Average Loss}}$$

$$MACD_t = EMA_{\text{fast}}(P_t) - EMA_{\text{slow}}(P_t)$$

$$Upper_t = MA_t + k \sigma_t, \quad Lower_t = MA_t - k \sigma_t$$

Support and resistance, trend strength, and volume trends are derived from these primitives and OHLCV. Signal validation uses a composite score,

$$Score(s_j) = \alpha_1 Relevance(s_j) + \alpha_2 Recency(s_j) + \alpha_3 Credibility(s_j),$$

which governs how retrieved evidence is filtered before integration into reports and recommendations (Wu et al., 13 Jul 2025).

4. Report generation, forecasting, and adaptation

The system’s data flow begins with ingestion of OHLCV data from exchanges, news from APIs, RSS feeds, and curated portals, social and KOL tracking, on-chain metrics, and signal indexing in the database. Preprocessing includes technical feature engineering over OHLCV, semantic annotation of news, and construction of a dynamic knowledge graph. Text is summarized and enriched through LLM processing and GraphRAG, while structured market data are transformed through indicator computation. Agent partial reports are then validated, scored, cached, integrated, and augmented to produce final financial reports and strategies (Wu et al., 13 Jul 2025).

The operational loop is explicitly described. A task $T$ is first classified and routed to agents A1–A4. Each agent constructs a retrieval query, pulls fresh data through RAG or Perplexity, validates signals with composite scoring, and produces a partial report $R_i$ that is cached with a TTL $\tau_i$. The partial reports are integrated into

$R_{raw} = f_{integrate}(R_1, R_2, R_3, R_4),$

then augmented through Perplexity retrieval to obtain

$EMA_t = \alpha P_t + (1 - \alpha) EMA_{t-1}$0

In parallel, a forecast module runs LLM-based and ML-based predictors and fuses them with adaptive weights. The final forecast is

$EMA_t = \alpha P_t + (1 - \alpha) EMA_{t-1}$1

The reflection module then ingests historical predictions, outcomes, cached reports, and user feedback, and updates fusion weights, retrieval scoring coefficients, prompt schemas, and temporal scopes (Wu et al., 13 Jul 2025).

The adaptive logic is formalized through rolling accuracy and directional correctness. Accuracy is defined as

$EMA_t = \alpha P_t + (1 - \alpha) EMA_{t-1}$2

and directional correctness, or win rate, as

$EMA_t = \alpha P_t + (1 - \alpha) EMA_{t-1}$3

The update rule is summarized as: if $EMA_t = \alpha P_t + (1 - \alpha) EMA_{t-1}$4, then $EMA_t = \alpha P_t + (1 - \alpha) EMA_{t-1}$5. This embeds historical performance directly into forecast fusion, so model trust is not fixed but adjusted according to observed outcomes.

Caching is likewise explicit: $EMA_t = \alpha P_t + (1 - \alpha) EMA_{t-1}$6 minutes for technical analysis, $EMA_t = \alpha P_t + (1 - \alpha) EMA_{t-1}$7 hours for market dynamics, and $EMA_t = \alpha P_t + (1 - \alpha) EMA_{t-1}$8 is dynamic for recommendations. If $EMA_t = \alpha P_t + (1 - \alpha) EMA_{t-1}$9, the corresponding report $$RSI = 100 - \frac{100}{1 + RS}, \quad \text{where } RS = \frac{\text{Average Gain}}{\text{Average Loss}}$$0 is reused. This policy is central to the system’s near-real-time behavior, balancing update frequency against the computational cost of a multi-agent retrieval pipeline (Wu et al., 13 Jul 2025).

5. Investment recommendations and user interaction

MountainLion generates multi-horizon recommendations that combine technical triggers with macroeconomic, regulatory, and capital-flow context. Report generation proceeds through task decomposition across A1–A4, parallel analysis, centralized integration, and Perplexity augmentation, with conflicts resolved by prioritizing higher-confidence sources. The user-facing interface supports presets such as comprehensive analysis, conversational AI chat, alerts, question answering, and iterative refinement of recommendations (Wu et al., 13 Jul 2025).

The example report structure is organized across short-, medium-, and long-term horizons. For the short term, defined as 1–4 weeks, technical triggers are enriched with on-chain wallet accumulation, liquidation volume spikes, and funding rate moves. For the medium term, defined as 1–6 months, reports incorporate ETF inflows, regulatory clarity in the U.S. and EU, policy easing, capital flows, and sentiment narratives supporting trend continuation or reversal. For the long term, defined as 6+ months, the emphasis shifts to institutional adoption and declining exchange reserves as structural supply signals underpinning allocation and DCA logic (Wu et al., 13 Jul 2025).

The paper argues that this structure improves practical applicability because recommendations are not presented as isolated indicator outputs. Instead, they are embedded in evidence trails and contextual narratives. A plausible implication is that the system aims to bridge discretionary and systematic trading styles by rendering machine-generated signals legible in terms commonly used by human market participants, especially when sentiment, policy developments, and capital flows materially interact with price dynamics.

6. Evaluation, interpretability, and limitations

The evaluation setup uses OHLCV data across multiple granularities and tokens, with forecasts archived in resolution-specific tables, together with news from APIs, RSS, and curated portals, and with on-chain metrics and social or KOL tracking (Wu et al., 13 Jul 2025). The baseline is a minimal technical-only strategy based on support and resistance and oscillator signals, without macro, on-chain, or ETF overlays. ChatGPT-4o, DeepSeek-V3, and Grok-3 are compared as LLM agents tasked with refining reports and recommendations using market signals and narratives.

The paper reports token-level results in a table titled “Forecasting Results Across Tokens,” where MSE and CV scores are reported across ADA, BTC, ARB, SOL, XRP, DOGE, TRX, ETH, MATIC, and BNB. TRX is reported as showing the best performance, while BTC and ETH exhibit higher MSE, which the paper attributes to large-cap nonstationarities. Comparative analysis further indicates that ChatGPT-4o emphasizes whale accumulation, ETF net inflow triggers, and dynamic volatility thresholds for partial profit-taking; DeepSeek-V3 emphasizes policy overlays, on-chain flows, and adaptive cash buffers; and Grok-3 prioritizes whale transfers, IMF classification signals, and momentum redeployment. The key qualitative outcome is that the medium-term window of 1–6 months is the horizon most improved by LLM-driven reports blending institutional narratives, policy shifts, and execution flexibility (Wu et al., 13 Jul 2025).

Interpretability is described as arising from evidence trails via the knowledge graph and retrieved sources, together with narrative coherence introduced by A4. Before-and-after refinements are reported qualitatively: short-term technical triggers gain credibility when supported by wallet increases and liquidation spikes; medium-term reports become more persuasive for institutional and mid-sized investors when they include ETF flows, regulatory clarity, and easing signals; and long-term theses are strengthened by institutional adoption and declining exchange reserves. The paper therefore links interpretability to evidentiary grounding and narrative integration rather than to transparent parametric structure alone.

Several limitations are identified. Data quality and timeliness, including news noise and on-chain data lags, may affect reasoning. Multi-agent and RAG pipelines remain computationally heavier than static DL approaches, even with caching. Generalization under regime shifts remains difficult, and large-cap nonstationarity increases forecasting difficulty. The paper also notes the lack of end-to-end backtests with risk-adjusted metrics and the absence of execution-friction and slippage evaluation (Wu et al., 13 Jul 2025). Standard definitions for cumulative return, annualized return, volatility, Sharpe, Sortino, max drawdown, Calmar, and Kelly sizing are listed only as future backtesting metrics rather than reported empirical outcomes.

7. Implementation status, reproducibility, and prospective extensions

MountainLion operates primarily through prompting, retrieval, graph reasoning, and tool use, and the paper states that no end-to-end fine-tuning pipeline is required for the core LLM agents (Wu et al., 13 Jul 2025). ChatGPT-4o, DeepSeek-V3, and Grok-3 were used to refine baseline strategies in the case study, while a LLAMA2 13B model was fine-tuned on Web3 industry data in a demonstration context to improve filtering and categorization efficiency. Retrieval and indexing rely on RAG and GraphRAG for semantic enrichment, a Perplexity-based retriever for contextual augmentation, and MySQL for time-series storage and embeddings. Real-time capability is supported by continuous ingestion, TTL-based caching, and a low-latency ML track complementing LLM forecasts.

The reproducibility material includes input specifications, such as OHLCV tuples $$RSI = 100 - \frac{100}{1 + RS}, \quad \text{where } RS = \frac{\text{Average Gain}}{\text{Average Loss}}$$1, aggregated sequences $$RSI = 100 - \frac{100}{1 + RS}, \quad \text{where } RS = \frac{\text{Average Gain}}{\text{Average Loss}}$$2 and $$RSI = 100 - \frac{100}{1 + RS}, \quad \text{where } RS = \frac{\text{Average Gain}}{\text{Average Loss}}$$3, sentiment embeddings from historical and real-time news, and outputs including LLM multi-step OHLCV predictions, ML next-step price vectors, fusion via $$RSI = 100 - \frac{100}{1 + RS}, \quad \text{where } RS = \frac{\text{Average Gain}}{\text{Average Loss}}$$4, win rate and accuracy metrics, and a template-conditioned text generator $$RSI = 100 - \frac{100}{1 + RS}, \quad \text{where } RS = \frac{\text{Average Gain}}{\text{Average Loss}}$$5 for readable forecast reports. Prompts and templates include short-, medium-, and long-term search windows, JSON outputs with source, snippet, time, and URL, and multilingual prompts for Perplexity. The paper provides detailed schemas, pseudocode-like UI routing in Vue.js, and pipeline descriptions, but it does not provide a public code repository link or dataset dumps (Wu et al., 13 Jul 2025).

Future work proposed in the paper includes explicit outcome-driven prompt tuning, multi-armed bandit selection of retrieval policies, integration of robust volatility and regime-switching models, extension to cross-asset macro hedging with DXY, Nasdaq, and gold correlations, addition of pixel-level chart understanding through ViT or CLIP if needed, improved multilingual event extraction, and a full backtesting suite with Sharpe, Sortino, Calmar, and statistical significance tests (Wu et al., 13 Jul 2025). These proposed directions suggest an unresolved tension between the current system’s emphasis on interpretability and adaptive orchestration and the still-limited quantitative validation of trading performance under realistic execution assumptions.