Joint LLM-Infused Embeddings for Explainable Financial Recommendations at Scale

Joint LLM-Infused Embeddings for Explainable Financial Recommendations at Scale — Ai4

Ai4

AI-assisted summary of Joint LLM-Infused Embeddings for Explainable Financial Recommendations at Scale — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by Ai4 on August 14, 2026, this talk features Alejandro Correa Bahnsen, VP of Data & AI at Grupo Bursátil Mexicano (GBM). According to the video description, Bahnsen presents an explainable financial recommender system that bridges investor behavioral signals and market dependency structures using fine-tuned LLM-based embeddings within a shared representation space.

## TL;DR Addressing the challenges of scaling Mexico's largest brokerage app to nearly two million digital users, Bahnsen details GBM's progression from manual CIO curation to collaborative filtering and handcrafted Learning-to-Rank pipelines. While tabular feature engineering raised sales conversion by nearly 6%, it created a representation bottleneck that struggled to capture complex portfolio diversification, price reflexivity, and nuanced investor behavior. To solve this, GBM engineered a joint semantic embedding architecture utilizing three dedicated LLM encoders for user behavior, portfolio holdings, and candidate assets. The joint representation resolved asset cold-start issues, doubled offline NDCG and recall performance compared to engineered baselines, and established the groundwork for real-time conversational and multimodal voice advisory.

## Atmosphere & style The presentation carries the pragmatic, methodical cadence of an experienced practitioner operating in a high-stakes financial environment. Bahnsen pairs transparent institutional war stories with architectural depth, balancing an analytical mindset with genuine engineering enthusiasm as he dissects why standard consumer recommendation tricks break down in wealth management. The delivery feels like an authoritative technical masterclass—grounded, measured, and focused on solving hard production constraints rather than chasing AI hype.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=4s) Financial advice is fundamentally subjective and depends on each investor's unique goals, time horizon, and risk tolerance rather than a universal right answer. - [00:04](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=4s) Scaling Grupo Bursátil Mexicano's digital investment platform to nearly two million users broke legacy manual curation models, which relied on the Chief Investment Officer selecting weekly picks. - [03:04](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=184s) Applying an initial collaborative filtering baseline drove a 3% uplift in sales, but treating investments like e-commerce purchases ignores wealth management fundamentals. - [05:31](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=331s) Financial recommendation systems differ fundamentally from e-commerce because they must prioritize portfolio diversification, risk preferences, and volatile market dynamics rather than mere engagement metrics. - [05:31](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=331s) Unlike consumer shopping, financial recommendations at scale can trigger price reflexivity, where algorithm-driven buying pressure directly moves asset prices in small-cap or domestic markets. - [11:41](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=701s) Adopting a classic Learning-to-Rank framework with handcrafted features across investor profiles, portfolio holdings, behavior, and market signals achieved a nearly 6% sales lift in production, doubling collaborative filtering. - [16:14](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=974s) Traditional tabular feature engineering hits a representation bottleneck where adding thousands of manual features produces diminishing returns and fails to capture rich financial context. - [18:35](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=1115s) A joint semantic embedding architecture replaces manual feature tables with three dedicated LLM encoders covering user behavior, portfolio holdings, and candidate assets. - [18:35](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=1115s) Joint LLM embeddings solve the cold start problem for newly listed IPOs and assets with limited trading history by leveraging foundation models' world knowledge and descriptive metadata. - [21:48](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=1308s) Offline benchmarking over six weeks revealed that joint semantic embeddings roughly doubled ranking accuracy across NDCG and Recall metrics compared to the feature-engineered baseline. - [23:58](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=1438s) Semantic representation learning lays the foundation for conversational investing, requiring strict portfolio guardrails, sub-second latency, and omnichannel consistency between chat interfaces and app UI feeds.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=4s) The Subjectivity of Financial Advice and GBM's Digital Transformation

Every day in asset management and financial advisory comes with the same recurring questions from friends, colleagues, and advisors alike: "Which stock is better? Should I buy Apple? Should I buy SpaceX?" The reality is that there is never a single universal answer. Recommendations depend fundamentally on who the investor is, what their specific goals are, their time horizon, and whether they are treating the platform as a place to gamble or as a home for their retirement savings.

Grupo Bursátil Mexicano (GBM) operates as an established, nearly 40-year-old financial advisory and brokerage firm. Over the past three to five years, however, the firm expanded rapidly into the digital space, scaling its investment app to close to two million digital users and cementing its position as by far the largest brokerage firm in Mexico, even while maintaining its traditional offline operations.

In the app's early digital days, the homepage presented identical content to all users. Every Monday morning, curation was handled manually: the Chief Investment Officer (CIO) would decide which investment opportunities or popular instruments to highlight for that week. While this static, editorial method worked passably when managing tens of thousands of clients, it completely broke down when scaling to millions of users with diverse risk profiles and financial needs.

```mermaid flowchart TD accTitle: Evolution of Recommendation Systems at GBM accDescr: Transition from manual CIO weekly picks to collaborative filtering and machine learning models.

subgraph Phase1["Legacy Manual Approach"] CIO["Chief Investment Officer (CIO)"] -->|"Manual weekly picks"| StaticFeed["Static Homepage Feed<br/>(Same for all users)"] end

subgraph Phase2["Initial Personalization"] UserActions["User Buy History"] --> CF["Collaborative Filtering"] CF -->|"Similar users bought X"| RecFeed["Personalized Feed<br/>(+3% conversion vs baseline)"] end

Phase1 -.->|"Fails at millions of users"| Phase2 ```

---

[03:04](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=184s) The First Step: Collaborative Filtering and Its Limits

To solve the scale challenge, the team looked toward standard e-commerce and media platforms that personalize user feeds based on inferred intent. Borrowing the classic collaborative filtering paradigm—the idea that "similar customers bought item X, therefore you should buy item X"—GBM deployed a collaborative filtering model to drive homepage recommendations.

Implemented at the beginning of the year, the collaborative filtering model delivered an immediate +3% uplift in investment sales compared to the baseline of showing the most popular assets. Initially, performance was tracked using a straightforward metric: comparing the amount of money invested by users who were exposed to the personalized recommendation module against those who were not.

- **Initial Metric:** Aggregate capital invested by users exposed to personalized recommendations vs. non-exposed users. - **Immediate Outcome:** +3% higher conversion/sales over the static popular-instrument baseline. - **Limitation:** Recommending complex assets solely based on peer transaction co-occurrence ignores fundamental differences between consumer shopping and wealth management.

---

[05:31](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=331s) Why Financial Recommendations Are Fundamentally Different from E-Commerce

While consumer platforms optimize for user attention, video watch time, retention, click-through rates, social engagement, or recurring subscription revenue, measuring financial recommendations is far more challenging. In consumer platforms, the core question is "What should I show next?"; in finance, the user is asking, "What should I do with my money?"

A sound financial recommendation system cannot simply optimize for immediate engagement. It must balance multiple competing dimensions: 1. **Uncertain Returns & Risk Preferences:** Users range from highly risk-averse to aggressive investors. Predicting returns is practically impossible; the only reliable signal is historical volatility. 2. **Portfolio Diversification:** If an investor's portfolio is already concentrated in technology equities, recommending yet another tech stock is irresponsible, even if their behavioral data indicates a strong preference for tech. 3. **Price Reflexivity (Market Impact):** Unlike consumer products, recommending an asset at scale to millions of users can directly alter market mechanics. For small-cap, niche, or domestic securities, routing millions of impressions to an asset creates artificial buying pressure that moves the underlying price—a phenomenon unique to finance that acts as an unintended byproduct of the algorithm. 4. **Temporal Decay and Market Dynamics:** The shelf life of a recommendation is tied to live market swings. Showing an asset whose price has just experienced extreme volatility changes the nature and safety of the recommendation.

| Dimension | Standard E-Commerce / Streaming | Financial Investment Platforms | |:--- |:--- |:--- | | **Primary User Goal** | Discover content/products to consume | Manage capital and build wealth safely | | **Core Optimization Metrics** | CTR, watch time, retention, GMV | Risk-adjusted return, diversification, volatility alignment | | **Asset Concentration Impact** | Low harm (e.g., listening to one genre) | Severe risk (over-indexing on a single sector) | | **Market Feedback Loop** | Item availability/price generally static | **Price reflexivity**: Recommendations can move asset prices | | **Temporal Sensitivity** | Moderate (decays over months/years) | Extreme (decays over hours/days based on price swings) |

---

[11:41](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=701s) Engineering Machine Learning Pipelines with Learning-to-Rank

To move from predicting "what will the user click" to "what is financially relevant for this specific investor right now," recommendation models must ingest richer, multi-dimensional user context. Financial relevance depends on an investor's existing holdings, transaction history, search and price tracking habits, position sizing adjustments (increasing vs. decreasing exposure), and specific fiscal constraints such as capital gains tax liabilities.

To capture these interactions, the engineering team's first robust modeling attempt centered on a classic Machine Learning **Learning-to-Rank (LTR)** framework. This required data science teams to spend weeks handcrafting domain-specific features across multiple data domains:

```mermaid flowchart LR accTitle: Feature Engineering Pipeline for Financial Learning-to-Rank accDescr: Multiple data sources combined into engineered feature sets fed into a Learning-to-Rank model.

subgraph DataSources["Data Sources & Context"] U["Investor Profile & Demographics"] P["Current Holdings & Tax Factors"] I["Interaction History (Search, Views)"] M["Market Signals & Asset Price Shifts"] end

subgraph FeatureEng["Handcrafted Feature Pipeline"] F1["User Risk & Balance Features"] F2["Portfolio Concentration Metrics"] F3["Behavioral Intent Signals"] F4["Dynamic Market Features (e.g., Price Drops)"] end

subgraph ModelStage["Ranking Layer"] LTR["Learning-to-Rank (LTR) Model"] Output["Ranked Investment Recommendations"] end

U --> F1 P --> F2 I --> F3 M --> F4

F1 & F2 & F3 & F4 --> LTR LTR --> Output ```

- **User & Investor Features:** Risk tolerance scores, investment horizons, account tier, and historical trading style. - **Portfolio Features:** Asset allocation breakdown, existing sector weights, and tax sensitivity parameters. - **Interaction & In-App Activity:** Search logs, watchlist queries, price chart interactions, and frequency of position scaling. - **Market & Asset Features:** Relative price movements, volatility windows, and trigger features (e.g., whether a user is viewing an asset that recently pulled back in price).

[14:47](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=887s) Production Impact: Doubling Recommender Performance

Deploying personalized recommendations in-app—such as curating top-five asset lists for carousel widgets—delivered an immediate commercial impact. A production rollout in March, evaluated over a month of live testing, revealed that personalized asset recommendations drove an increase in sales of nearly 6% compared to merely displaying the most popular assets. Across a user base scaling into millions of investors, this uplift represents substantial trading volume.

The result is particularly striking given the architectural simplicity of the underlying machine learning model, where the bulk of the effort was focused heavily on manual feature engineering. Compared to the baseline collaborative filtering model, which achieved a 3% sales lift, this iteration successfully doubled performance.

- **Production Uplift:** Almost 6% increase in sales over showing baseline popular assets. - **Baseline Comparison:** Doubled the ~3% sales improvement previously achieved via collaborative filtering. - **Application Context:** Powered personalized in-app carousels displaying top-5 asset recommendations across millions of active users.

```mermaid flowchart LR accTitle: Recommender Model Iteration Performance subgraph Baselines ["Baseline Models"] A["Popular Assets Baseline"] -->|Baseline| B["0% Relative Lift"] C["Collaborative Filtering"] -->|Initial ML| D["+3% Sales Lift"] end subgraph Engineered ["Engineered Model"] E["Feature-Engineered Ranker"] -->|Doubled Impact| F["+6% Sales Lift"] end ```

---

[16:14](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=974s) The Representation Gap and Feature Engineering Bottlenecks

Despite the strong production metrics, the team identified severe structural limitations inherent to traditional tabular feature engineering, termed the "representation gap." Manually engineered features fail to capture the nuanced, high-dimensional reality of financial assets and dynamic investor behavior. On the asset side, analysts track industry sectors (e.g., consumer technology), recent revenue trajectories, corporate buybacks, and shifting fundamentals. On the user side, signals multiply rapidly: recent transaction timing, whether an investor panic-sells immediately upon a price dip or waits out drawdowns, and how they consume in-app financial news—including whether the sentiment of the articles they read about a particular holding is positive or negative.

Attempting to hand-craft features for every signal quickly hits diminishing returns. Expanding feature sets from 10,000 to 11,000 tabular inputs stopped yielding performance gains and merely introduced noise, proving that "not every interaction can be anticipated and manually created into a feature." This realization prompted a strategic shift away from manual financial feature engineering toward learning rich, dense semantic representations.

---

[18:35](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=1115s) Architecture: Joint LLM-Infused Semantic Embeddings

To bridge the representation gap, the team developed a joint semantic embedding architecture that replaces manual feature tables with three dedicated encoders powered by large language models. The framework simultaneously encodes three core pillars of financial context: the user's granular recent behavior, their current portfolio composition, and the universe of candidate assets available on the platform.

```mermaid flowchart TD accTitle: Joint Semantic Recommender Architecture subgraph Inputs ["Input Encoders"] U["User Recent Behavior"] --> E1["LLM Behavior Encoder"] P["User Portfolio Holdings"] --> E2["LLM Portfolio Encoder"] A["Candidate Asset Universe"] --> E3["LLM Asset Encoder"] end subgraph Fusion ["Fusion & Scoring"] E1 --> C["Concatenation Layer"] E2 --> C E3 --> C C --> NN["Neural Network Ranker"] NN --> Out["Relevance Score / Top-K Assets"] end ```

The resulting embeddings from all three branches are concatenated into a unified representation space and fed into a downstream neural network to generate a personalized relevance score. This setup allows the system to harness the inherent world knowledge of foundation models regarding public companies and macroeconomic relationships. Crucially, semantic embeddings also resolve the severe "cold start problem in financial recommendations"—allowing newly listed assets, such as recent high-profile IPOs that lack historical trading interactions, to be effectively embedded and recommended based on descriptive metadata without manual tabular intervention.

---

[21:48](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=1308s) Offline Validation: Doubling Ranking Accuracy

Offline evaluation of the new joint representation ranker demonstrated dramatic improvements across standard ranking metrics when benchmarked against previous feature-engineered iterations. Testing across offline historical datasets showed that semantic representations roughly doubled performance on Normalized Discounted Cumulative Gain (NDCG) as well as Recall metrics, accurately mirroring what users subsequently purchased.

While acknowledging a personal preference to present live production metrics, the offline experiments over the preceding six weeks provided strong empirical justification for an imminent production rollout. The findings confirm that encoding user portfolios, behavioral journeys, and asset metadata into a shared semantic space captures far more predictive investor intent than traditional feature tables can express.

- **Offline Gains:** Approximately 2x improvement in NDCG and Recall compared to the feature-engineered production baseline. - **Validation Scope:** Six weeks of experimental offline benchmarking prior to scheduling live production deployment. - **Core Objective:** Accurately replicating human financial advisor intuition within automated digital experiences.

---

[23:58](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=1438s) Conversational Investing, Multimodal Voice, and Guardrails

The long-term value of joint semantic modeling lies in its natural compatibility with the broader shift toward conversational finance, encompassing interactive text chat and multimodal voice interfaces. Traditional recommendation tables cannot adapt dynamically to an ongoing dialogue; in contrast, a semantic encoder can ingest conversational turns in real time and immediately refresh portfolio recommendations.

Operating financial AI in conversational channels introduces strict technical and regulatory challenges:

- **Portfolio Guardrails:** Preventing overexposure to specific sectors or volatile asset classes when an investor’s risk profile dictates diversification. - **Omnichannel Consistency:** Ensuring that recommendations provided inside a conversational chat or voice session match the assets displayed on the app’s homepage carousel, rather than relying on a disconnected generic LLM that hallucinates investment advice. - **Sub-Second Latency:** Transitioning semantic encoding pipelines to real-time execution speeds necessary to support seamless, live conversational interactions with digital financial advisors.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=4s) Which stock is better? Should I buy Apple? Should I buy SpaceX? — Alejandro Correa Bahnsen

> [03:04](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=184s) similar customers bought item X, therefore you should buy item X — Alejandro Correa Bahnsen

> [05:31](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=331s) What should I show next? — Alejandro Correa Bahnsen

> [05:31](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=331s) What should I do with my money? — Alejandro Correa Bahnsen

> [11:41](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=701s) what will the user click — Alejandro Correa Bahnsen

> [11:41](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=701s) what is financially relevant for this specific investor right now — Alejandro Correa Bahnsen

> [16:14](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=974s) representation gap — Alejandro Correa Bahnsen

> [16:14](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=974s) not every interaction can be anticipated and manually created into a feature — Alejandro Correa Bahnsen

> [18:35](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=1115s) cold start problem in financial recommendations — Alejandro Correa Bahnsen

## Mentioned Alejandro Correa Bahnsen, Grupo Bursátil Mexicano (GBM), Apple, SpaceX, Normalized Discounted Cumulative Gain (NDCG), Learning-to-Rank (LTR)

## Video description Alejandro Correa Bahnsen - VP of Data & AI, Grupo Bursátil Mexicano

Traditional recommendation systems such as collaborative filtering and market-signal models often fail to capture investor intent and rarely provide meaningful explanations. In this talk, I will present a financial recommender that jointly models investor behavioral signals (holdings, watchlists, interaction patterns) and market dependency structures using fine-tuned LLM-based embeddings in a shared representation space.

I will explain how this joint embedding approach improves relevance, diversity, and stability compared to behavior-only or market-only baselines, and how we evaluate these gains both offline and in production-oriented settings. The talk will also cover architectural decisions, representation learning trade-offs, and explainability strategies needed for real-world deployment.

If you are building ranking systems in high-stakes domains, this session will offer concrete patterns for combining foundation models with structured financial data in a scalable and interpretable way.

[Source](https://www.youtube.com/watch?v=Kgp27I4VNYw)

中文简报

融合LLM联合嵌入的规模化可解释金融推荐 为应对将墨西哥最大经纪商应用扩展至近两百万数字用户的挑战，Bahnsen详细介绍了GBM如何从首席投资官（CIO）人工精选逐步演进到协同过滤及手工构建的排序学习流程。尽管表格特征工程使销售转化率提升了近6%，但它带来了表征瓶颈，难以捕捉复杂的投资组合分散化、价格反身性以及细微的投资者行为。为此，GBM设计了一种联合语义嵌入架构，利用三个专用的LLM编码器分别处理用户行为、投资组合持仓及候选资产。该联合表征解决了资产冷启动问题，相比特征工程基线使离线NDCG与召回率翻倍，并为实时对话式及多模态语音投顾服务奠定了基础。 财务建议本质上是主观的，取决于每位投资者独特的理财目标、投资期限和风险承受能力，并不存在放之四海而皆准的标准答案。 (0:04) 墨西哥证券经纪集团（Grupo Bursátil Mexicano）的数字化投资平台拓展至近两百万用户，突破了以往依赖首席投资官每周挑选标的的传统人工精选模式。 (0:04) 最初应用协同过滤基线带来了3%的销售提升，但将投资视同电商购物忽略了财富管理的基本逻辑。 (3:04) 金融推荐系统与电子商务有着本质区别，因为它必须优先考量资产组合分散度、风险偏好以及剧烈的市场动态，而不仅是单纯的互动指标。 (5:31) 与普通消费购物不同，规模化金融推荐可能会引发价格反身性，即算法驱动的买盘压力会直接影响小盘股或本土市场的资产价格。 (5:31)

Reading summary

To scale personalized financial advice to millions of digital investors, Mexican brokerage Grupo Bursátil Mexicano (GBM) transitioned from manual feature engineering to a joint LLM-infused semantic embedding architecture, roughly doubling recommender accuracy in offline benchmarks. Why Financial Recommendations Differ from E-Commerce (5:31) Standard recommendation engines optimize for consumer engagement metrics such as click-through rates, watch time, and immediate purchase conversion. Financial advisory systems, however, cannot treat asset selection like consumer goods shopping. Recommending assets requires balancing personal risk tolerance against volatile market returns, enforcing portfolio diversification to prevent sector concentration, and accounting for temporal price decay. Furthermore, financial platforms must account for price reflexivity: directing millions of digital users toward smaller domestic or niche securities creates artificial buy-side pressure that can distort the underlying market price. The Evolution from Collaborative Filtering to Learning-to-Rank (11:41) GBM historically handled homepage curation manually through its Chief Investment Officer's weekly picks, a method that failed as the user base expanded toward two million. Deploying an e-commerce-style collaborative filtering model based on peer purchase history produced an initial 3% uplift in investment sales over showing popular assets. To address the complexities of investor holdings and tax constraints, the engineering team next deployed a machine learning Learning-to-Rank (LTR) framework that combined handcrafted user risk metrics, portfolio allocations, in-app behavioral signals, and asset price fluctuations. In a month-long production test during March, this feature-engineered ranker generated an almost 6% increase in sales over the popular-asset baseline, doubling the lift achieved by collaborative filtering. The Representation Gap in Tabular Feature Engineering (16:14) Despite strong commercial results powering top-five asset carousels, the team encountered a representation bottleneck. Handcrafted tabular features could not adequately capture high-dimensional signals, such as how an investor responds to portfolio drawdowns, their news sentiment consumption, or changing corporate fundamentals. Attempting to expand tabular inputs from 10,000 to 11,000 features yielded diminishing returns and introduced excess noise, proving that dynamic investor intent cannot be fully modeled through manual feature generation alone. Joint LLM Embeddings and Offline Performance Gains (18:35) GBM engineered an architecture featuring three dedicated large language model encoders that capture user behavior, current portfolio allocations, and available platform assets. These dense semantic embeddings are concatenated into a shared representation space and scored by a downstream neural network ranker. In offline evaluations conducted over six weeks, this semantic ranker roughly doubled performance on Normalized Discounted Cumulative Gain (NDCG) and Recall metrics relative to the production feature-engineered baseline. The model also addresses the cold-start problem for newly listed assets like initial public offerings, using descriptive metadata rather than historical trading interactions. Deployment Constraints and Conversational Integration (23:58) The architecture is designed to support conversational text and multimodal voice interfaces by updating recommendations dynamically as a user speaks or chats. However, conversational financial advisory introduces strict operational guardrails. Systems must enforce portfolio diversification limits to prevent overexposure, maintain omnichannel consistency so that conversational suggestions align precisely with homepage carousel recommendations, and maintain sub-second latency for live interactions. While joint semantic embeddings roughly doubled ranking metrics in six weeks of offline historical validation, live production metrics remain pending, and real-time deployment depends on maintaining low latency and strict diversification guardrails.

精炼阅读摘要

针对传统电商推荐逻辑与人工表格特征工程在财富管理场景中的失效问题，墨西哥头部券商 GBM 提出融合大语言模型（LLM）的多路联合语义嵌入架构，有效克服了金融推荐中的组合分散限制、表征鸿沟与新股冷启动障碍。 金融推荐与传统电商的本质分歧 (5:31) GBM 在从近40年历史的传统经纪业务向拥有近200万数字用户的平台转型过程中，早期依赖首席投资官（CIO）每周一手动挑选资产置顶，该方法在用户规模扩大后难以为继。团队随后引入电商常用的协同过滤算法，通过分析类似用户的购买记录推荐资产，在年初带来了相较于基线热门资产推荐约 3% 的投资销售额提升。然而，团队指出金融资产配置与电商消费品推荐存在本质差异：电商优化的是点击率、观看时长和商品交易额，用户关心“接下来看什么”；金融平台面临的核心诉求则是“如何处置资金”，直接关乎财富安全。 金融推荐系统必须平衡多重约束：其一，收益高度不确定且难以预测，系统只能将历史波动率作为可靠的风险信号；其二，必须强制兼顾资产组合的分散化，即便数据表明用户对科技股有强烈偏好，继续向其科技重仓组合推荐同类资产仍属不负责任；其三是金融特有的价格反身性，向数百万用户推荐中小市值资产会形成人为买压并推高市场价格，产生算法外生扰动；最后，金融资产存在极强的时效衰减，市场短期剧烈波动会迅速改变推荐资产的安全性与性质。 特征排序落地突破与表征工程瓶颈 (11:41) 为准确捕捉投资者的真实财务相关性，GBM 工程团队构建了基于机器学习的学习排序（LTR）模型。数据科学团队耗费数周手工构建了跨维度特征体系：涵盖用户风险偏好与交易风格的画像特征、行业权重及资本利得税敏感度等组合特征、搜索及自选浏览等交互信号，以及价格回撤与波动窗口等市场动态特征。在3月份开展的为期一个月的线上生产测试中，应用于应用首页 Top-5 资产轮播推荐位的排序模型实现了近 6% 的销售额增长，相比基线提升明显，且直接将此前协同过滤模型的 3% 增幅翻倍。 尽管商业指标显著改善，团队随后遇到了被称为“表征鸿沟”的结构性瓶颈。手工表格特征无法有效捕捉高维多变的金融现实，例如投资者在资产回撤时是恐慌抛售还是耐受持有、阅读的持仓新闻情绪是正面还是负面等复杂行为。当特征工程从 10,000 个特征强行扩展至 11,000 个时，模型性能不再提升并引入噪声，表明依靠人工穷举设计特征的方式已触及边际收益递减的上限，促使团队转向稠密语义表征学习。 融合大模型的多路联合语义表征 (18:35) 为了弥合表征鸿沟，团队设计了联合语义嵌入架构，用三个由大语言模型驱动的并行编码器取代手工表格特征：分别提取用户的近期细粒度行为、当前投资组合构成，以及平台候选资产全集的上下文语义。三组嵌入向量在融合层进行拼接，并输入下游神经网络排序模型以计算个性化相关性得分，从而输出推荐列表。 该架构不仅借助了通用基础模型对上市公司业务和宏观经济联系的既有认知，还从根本上解决了金融推荐中的严重冷启动难题。对于刚刚挂牌上市、缺乏历史交易和行为数据的高关注度首次公开募股（IPO）资产，系统无需依赖手工表格特征干预，即可直接依据描述性元数据生成高质量向量表征，实现向适宜用户的精准推荐。 离线性能翻倍与会话投顾演进约束 (21:48) 在连续六周的历史离线数据评估中，联合语义表征模型展现出大幅优于传统特征工程基线的排序表现，在归一化折损累计增益（NDCG）和召回率（Recall）指标上均实现了大约两倍的提升，精确拟合了用户随后的实际交易选择。团队表示，这一离线评测为其后续投入实际生产部署提供了坚实的经验依据，证明语义嵌入能更有效地映射人类专业投资顾问的直觉判断。 团队认为，联合语义模型的长远价值在于适配文本聊天与多模态语音等会话式投资界面的演变。语义编码器支持在多轮交互中实时刷新推荐组合，但面向实际业务落地仍须解决三大工程与合规挑战：一是建立组合安全护栏，防止风险厌恶型用户被诱导过度集中于高风险标的；二是保持全渠道推荐一致性，确保对话与 App 首页展示结果同源，杜绝通用 LLM 产生虚假投资建议；三是实现亚秒级推理延迟，以满足实时对话式数字投顾的交互需求。 该架构目前的核心成效基于为期六周的离线评测数据，其在数百万真实用户生产环境中的全量表现与价格反身性反馈仍有待上线后进一步检验；同时，如何在保障亚秒级低延迟响应的同时，严格实施组合风控护栏并消除全渠道交互中的大模型幻觉，构成了迈向会话式投资界面的关键现实制约。

Detailed notes — zh

## Background 本内容由墨西哥大型券商 Grupo Bursátil Mexicano（GBM）的数据与 AI 副总裁 Alejandro Correa Bahnsen 于 2026 年 8 月 14 日在 YouTube 平台的 Ai4 频道发布。根据页面简介，演讲重点探讨了如何在大规模金融场景中打破传统协同过滤与市场信号模型的局限，利用微调的大语言模型（LLM）将投资者行为信号与市场依赖结构融合在共享表征空间中，实现高相关性、多样性且兼具可解释性的智能金融资产推荐。 ## TL;DR 随着券商数字化用户突破数百万，GBM 经历了从 CIO 每周人工选品、传统协同过滤到基于排序学习（LTR）特征工程的推荐系统演进，线上测试分别带动了 3% 和近 6% 的投资销售额增长。然而，面对金融推荐特有的价格反身性、剧烈波动时效性以及人工特征工程突破万维后的边际递减与“表征差距”，团队转向了基于 LLM 的联合语义表征架构。新系统分别对用户近期行为、持仓组合与候选标的资产进行三路语义编码并融合排序，不仅在离线评测中将 Recall 与 NDCG 翻倍，更有效攻克了 IPO 新股等冷启动难题。未来，团队正致力于推动端到端毫秒级实时推理，并在严格的风控与合规护栏下，为下一代对话式和语音数字投顾提供全站一致的中央推荐支撑。 ## Atmosphere & style 整场演讲呈现出浓厚的技术务实主义与金融工程严谨度，演讲者以资深算法负责人的从容姿态，抽丝剥茧地复盘了一家老牌券商在数百万用户体量下的数字化转型阵痛。现场叙事褪去了浮夸的 AI 概念炒作，既有对 A/B 测试业务增益的兴奋展现，也有对离线数据翻倍时未完成实盘检验的坦诚与克制。听感专业、沉着且逻辑致密，将宏观金融顾问的同理心与高精尖的多路深度语义网络无缝交织，带给听众一种在金融风控高压线下稳健探索系统演进的技术现场感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=4s) 金融推荐不存在标准答案，投资决策高度依赖于用户的风险承受力、投资目标与持有周期等个体异质性，传统的静态千人一面展示无法满足规模化需求。 - [03:39](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=219s) 协同过滤虽然在线上初期将投资销售额提升了 3%，但在金融场景下单纯以成交额为目标会导致严重的风险收益错配并破坏资产多样化配置。 - [08:32](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=512s) 金融推荐具备三大特有属性：对小众资产存在直接扰动价格的“价格反身性”、对行情剧烈震荡极度敏感的“时效性”，以及需从“点击偏好”转向建模“投资者深度意图”。 - [14:47](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=887s) 采用排序学习（Learning to Rank, LTR）进行手工特征工程后，线上 A/B 测试推动资产销售额提升近 6%，但很快遭遇无法捕捉非结构化动态信息的“表征差距”瓶颈。 - [16:20](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=980s) 当人工工程特征突破万维后边际效益递减并引入大量噪声，促使团队转向利用大语言模型抽取深层语义表征的新架构。 - [18:35](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=1115s) 联合表征架构通过三个 LLM 语义编码器分别抽取用户行为、投资组合与候选资产的嵌入向量，拼接后经深度神经网络输出个性化相关性得分。 - [20:15](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=1215s) 基于大语言模型先验知识的语义表征有效攻克了金融推荐的冷启动难题，使无历史交易记录的 IPO 新股等资产也能获得精准推荐。 - [21:48](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=1308s) 在离线评测中，联合表征排序器在 Recall 和 NDCG 等关键指标上的表现达到了上一代特征工程生产模型的两倍左右。 - [23:58](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=1438s) 面向对话与语音交互的未来趋势，推荐系统必须向毫秒级实时化演进，以根据动态对话上下文即时调整资产推荐。 - [25:00](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=1500s) 对话助手必须通过调用底层中央推荐引擎来确保全站一致性与合规风控，严禁通用 LLM 自由幻觉生成未经检验的投资建议。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=4s) 金融推荐的本质挑战与传统券商的数字化转型 在财富管理和投资顾问领域，从业者每天都会面对诸如“哪只股票更好？”“我该买苹果还是 SpaceX？”之类的问题。但现实中这类问题根本没有标准答案。投资决策高度依赖于用户的个人画像：投资目标是什么、持有期限多长、是寻求退休养老规划还是纯粹投机交易。因此，推荐不仅取决于资产本身的属性，更取决于投资者是谁。 演讲者所在的墨西哥大型券商 Grupo Bursátil Mexicano（GBM）拥有近 40 年的历史。在过去三到五年中，公司大力发展数字化应用，迅速积累了近 200 万数字用户，成为墨西哥规模最大的线上券商，同时在传统线下经纪业务中继续保持领先地位。在数字化转型初期，其 App 首页采取了最原始的“千人一面”静态展示方式，由首席投资官（CIO）在每周一早晨人工挑选当周要推广的“投资机会”或“热门标的”。这种人工干预的模式在数万客户规模下尚可运转，但随着用户体量激增到数百万级，人工挑选推荐已完全无法满足个性化需求。 ```mermaid flowchart LR accTitle: 传统推荐到个性化推荐的演进路径 subgraph S1["传统运营模式"] A["CIO 每周人工选品"] --> B["全量用户统一展示<br/>(静态千人一面)"] end subgraph S2["规模化瓶颈"] B --> C{"用户规模突破百万"} end subgraph S3["算法化探索"] C --> D["协同过滤 (CF)"] D --> E["Learning to Rank (LTR)"] E --> F["多模态/联合表征探索"] end ``` ### [03:39](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=219s) 协同过滤的初步成效及其在金融场景的局限性 为了解决百万级用户的个性化分发问题，团队借鉴了成熟电商平台的做法，率先落地了基于**协同过滤（Collaborative Filtering）**的经典算法，即“买了该资产的相似用户还买了什么”。虽然数学逻辑简单，但在上线初期就取得了显著效果：相比展示热门推荐的对照组，接受协同过滤个性化推荐的用户在投资销售额上提升了 **3%**。 然而，评估指标与平台机制的深入分析揭示了电商推荐与金融推荐之间的本质鸿沟。电商或流媒体平台通常围绕用户时长、留存率、点击率、社交互动或订阅收入进行优化，这些指标在数字消费场景中直观且易于追踪。但在金融推荐场景中，仅以“促成了多少投资金额”作为评估标准存在严重偏差： - **收益与风险错配**：资产预期收益本身无法精准预测，模型最多只能估算资产波动率。不同投资者的风险承受能力差异巨大，纯推荐无法平衡高风险偏好与保守型需求。 - **组合多样化与再平衡**：若投资者已重仓科技股，推荐系统不能盲目因为协同过滤评分高而继续推荐科技股，而应引导用户进行资产多元化配置。 - **角色定位与业务逻辑**：虽然 App 首页无法完全替代专业理财规划师，但算法逻辑必须内嵌理财顾问的核心原则，而非单纯追求促成交易。 | 评估维度 | 传统电商 / 流媒体推荐 | 金融资产推荐 | |:--- |:--- |:--- | | **核心问题** | "接下来应该向用户展示/播放什么？" | "用户应该如何配置/打理自己的资金？" | | **核心优化目标** | 停留时长、点击率 (CTR)、留存率、社交互动、转化率 | 风险收益匹配度、投资组合多样化、波动率控制、税收优化 | | **对资产价格的影响** | 无（展示商品不会直接影响全网市场价格） | 存在**价格反身性 (Price Reflexivity)**，小盘/本地资产受推荐冲击大 | | **时效性与衰减** | 偏向新鲜内容，但长尾历史内容仍稳定有效 | 价格高频波动，时效敏感度极高，存在动态风险 | ### [08:32](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=512s) 金融推荐的三大核心异质性：反身性、时效性与复杂意图 演讲者指出，金融产品推荐与普通消费品推荐存在三大底层差异： 1. **价格反身性（Price Reflexivity）**：在大规模金融推荐中，仅仅“推荐某个资产”这一动作本身，就可能直接驱动资产价格波动。对于苹果、谷歌等巨盘股影响有限，但对于小盘股或新兴市场的本地小众资产，向数百万用户集中曝光会导致买盘骤增、人为抬高价格。这种市场反馈是金融推荐独有的伴生效应。 2. **时效性与价格波动（Temporal Decay & Volatility）**：时效性在流媒体中多表现为推陈出新，而在金融领域中，资产价格的剧烈波动会彻底改变推荐的合理性。当某一标的处于剧烈震荡期时，系统必须动态决定是否继续向特定用户展示。 3. **从“点击偏好”转向“投资者深度意图”**：金融推荐不能仅以“用户会点击什么”为导向，而必须建模“该投资者在当前市场环境下最需要什么”。 为了构建真正有效的金融推荐，系统必须整合多维度的异构数据： - **组合与持仓结构**：现有持仓分布、现金比例及集中度风险。 - **交易与行为轨迹**：历史买卖记录、加仓与减仓操作、自选股列表（Watchlist）、高频查价与行情浏览行为。 - **财务与合规约束**：资产潜在资本利得税影响及投资者税收偏好。 ### [13:27](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=807s) 经典机器学习阶段：基于排序学习（Learning to Rank）的特征工程 为融合上述复杂特征，团队的首个进阶方案是将推荐重构为标准机器学习问题，采用**排序学习（Learning to Rank, LTR）**框架。在此架构下，数据科学家耗费数周时间手动构建了海量的特征工程流水线，主要覆盖四大维度： - **用户特征（User Features）**：风险等级、投资年限、资金规模等画像信息。 - **持仓特征（Portfolio Features）**：当前组合分布、资产类别暴露、集中度指标。 - **资产端特征（Asset Features）**：候选金融工具的行情表现、波动率、资产类型及近期涨跌幅。 - **交互历史特征（Interaction Features）**：用户在 App 内的搜索、点击、行情浏览以及“资产价格刚下跌时的浏览行为”等衍生交叉特征。 尽管这种手工构建特征并输入通用机器学习模型的方案具备可行性，但在面对海量非结构化市场信息与高动态变化时，人工特征工程的边际成本与表征瓶颈迅速显现。 ### [14:47](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=887s) 生产环境测试结果与“表征差距”的瓶颈 在实际落地推荐系统时，展示形式会直接取决于应用界面的具体模块。最典型的场景是在移动端 App 的信息轮播卡片（Carousel）中为每位用户呈现个性化的 Top-5 资产推荐。团队在 3 月份左右将特征工程驱动的推荐模型推向生产环境，经过大约一个月的线上 A/B 测试，数据反馈极为亮眼：相比于单纯向用户展示市场上最热门的头部资产，这套推荐模型直接推动资产销售额提升了将近 6%。考虑到平台拥有数以百万计的庞大用户体量，能够撬动 6% 的实际销售增长展现出了惊人的业务价值；相较于最初仅带来 3% 增幅的传统协同过滤（Collaborative Filtering）基线，这一轮以特征工程为核心的机器学习模型直接让业务回报翻了一倍。 然而，团队在兴奋之余迅速把精力投入到对模型局限性的复盘中。即便取得了不错的效果，团队依然敏锐地觉察到现有模型无法逾越的深层障碍，他们将其定义为“表征差距”（Representation Gap）。在传统的特征工程范式下，工程师必须手动将各种业务信号编码为特征输入模型，但海量非结构化和动态交互信息在这个过程中被严重损耗或曲解。 以分析苹果公司（Apple）这样单一资产为例，相关信息量是极其浩瀚的：从基础的消费电子行业归属、近期的营收财报数据，到公司的股票回购计划等基础面指标层出不穷。而在用户端，复杂性更是呈指数级上升：用户最近购买了什么？当资产价格跳水时，他们是恐慌性秒抛，还是选择承受亏损并在低位持仓观望？此外，平台还在 App 内嵌入了实时财经新闻模块，不仅能捕捉到用户正在阅读哪篇报道，还能推断该新闻对特定资产的情绪倾向是正面还是负面。如果单纯依赖传统机器学习，拼命向模型堆砌人工特征，很快就会遭遇天花板——当特征量从 10,000 个激增到 11,000 个时，模型几乎不再吸收有效信息，反而引入了大量破坏泛化能力的噪声。团队清醒地意识到，必须跳出手动金融工程的固有思维，转向能够捕获深层语义表征（Semantic Representations）的全新架构。 - **线上测试增益**：相比展示热门资产，新模型拉动了近 6% 的销售额增长，相比协同过滤的 3% 实现了翻倍。 - **表征差距痛点**：手动特征工程无法有效整合非结构化动态数据（如新闻舆情情绪、波动下的持仓心理）。 - **特征维度边际效益递减**：特征规模突破万维后，新加维度几乎只带来噪声，无法继续提升模型表现。 --- ### [18:35](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=1115s) 基于大语言模型的联合语义表征架构与冷启动破解 为了弥补这一表征鸿沟，团队设计了一套全新的联合表征框架。该方案不再依赖耗时费力的人工特征提取，而是借助标准大语言模型（LLM）构建了三个专门的语义编码器（Encoders），从三个核心维度全面抽取高维语义嵌入：第一是用户的近期行为（Recent Behavior），第二是用户的实际投资组合（Portfolio），第三是所有可选的目标候选资产（Candidate Assets）。三个编码器输出的嵌入向量随后被拼接（Concatenate）在一起，送入下游的神经网络进行交互融合，并最终输出目标资产对该用户的个性化相关性得分（Relevance Score）。 ```mermaid flowchart TD accTitle: Joint LLM-Infused Semantic Representation Recommender subgraph Inputs["多源输入信号"] U["用户近期行为数据"] P["用户当前持仓组合"] A["候选标的资产详情"] end subgraph Encoders["LLM 语义表征层"] E1["行为语义编码器"] E2["投资组合编码器"] E3["资产语义编码器"] end subgraph Fusion["交互排序层"] C["向量拼接 (Concatenate)"] NN["深度神经网络排序模型"] S["最终资产相关性得分 (Score)"] end U --> E1 P --> E2 A --> E3 E1 --> C E2 --> C E3 --> C C --> NN NN --> S ``` 这套架构的最大精妙之处在于“借力打力”：预训练大语言模型本身就在预训练阶段吸收了跨越互联网的广博金融知识，对各家上市公司的业务版图和行业逻辑有着天然的理解。将海量文本、新闻与动态交易模式原汁原味地喂给 LLM，既规避了机械式特征抽取的繁复工作，又赋予了下游排序网络极为丰沛的上下文语义。 在实际业务场景中，这套系统还奇迹般地攻克了金融推荐领域长期存在的“冷启动难题”（Cold Start Problem）。在传统协同过滤或行为模型中，如果某项资产此前从未被任何用户买入过——例如平台前一个月刚上线的重大 IPO 新股——系统就会因缺乏交互记录而彻底瘫痪，过去只能依赖人工干预编写规则强推。但在语义表征范式下，即使标的资产交互记录为零，LLM 依然能够通过招股说明、公司介绍和行业定位提炼出精准的语义特征向量，从而使新上市资产在冷启动阶段也能被合理排序与推荐。 - **三路编码输入**：用户行为（Behavior）、投资组合（Portfolio）、标的资产（Assets）。 - **流程精简**：LLM 抽取嵌入 $\to$ 向量拼接 $\to$ 神经网络计算相关度打分。 - **破解冷启动**：对无历史交易记录的新上市资产（如 IPO 新股），依靠文本与语义先验即可实现精准推荐。 --- ### [21:48](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=1308s) 离线评测翻倍与对标专业投顾经验 在投入线上真实流量之前，团队基于过去一个半月的离线实验对这一被命名为“联合表征排序器”（Joint Representation Ranker）的新算法展开了严苛的基准评测。评估主要围绕两个推荐系统关键指标：召回率（Recall）以及归一化折损累计增益（NDCG）。离线评测的数据曲线给团队注入了极强的信心——在预测用户最终实际购买标的资产的命中率上，联合表征排序器展现出的预测能力直接达到了上一代最佳特征工程模型的两倍左右。 尽管离线指标呈现出颠覆性的跃升，演讲者依旧保持着严谨求实的工程态度，坦言自己更希望在台上面向同行展示的是大规模真实资金检验后的线上 A/B 测试数据，而这项新技术预计在接下来的数周至一个月内才会正式推向生产系统。然而，离线实验的显著优势已充分确立了一个关键事实：全面采用语义表征来编码用户全貌与资产画像，其潜力远远超越了过去的孤立建模路径。 与此同时，团队在探索中也意识到技术依然处于起步阶段，当前的推荐逻辑相比现实世界中资深线下投资顾问（Financial Advisor）的服务水准仍有不小差距。真人理财顾问在面对客户时，往往兼具同理心、宏观市场认知与灵活的沟通艺术，能够挖掘出冰冷交易日志背后隐秘的财务愿望与风险焦虑。团队下一步的核心目标，正是尝试将更多这种顾问式的人性化洞察拆解、转化为数字体验，并持续沉淀到推荐架构中去。 - **离线核心评估指标**：Recall（衡量能否击中用户实际购买）与 NDCG（衡量推荐位排序质量）。 - **实验性能飞跃**：联合表征排序器的离线表现达到上一代生产模型的两倍。 - **数字化演进方向**：系统正逐步吸收并模拟线下专业投资顾问的资产配置与交互逻辑。 --- ### [23:58](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=1438s) 面向对话式投顾、语音交互与实时风控护栏 在展望推荐系统的技术终局时，演讲者强调整个数字金融交互界面正在不可逆转地向对话式体验（Conversational Experience）演进，无论是文本聊天还是多模态语音（Voice）形态。在此背景下，前几代基于批处理或静态表格特征的推荐模型在逻辑上完全失效——当用户正在与系统文字交流或进行语音问答时，后台必须根据对话流的实时推移立刻修正推荐列表，传统模型根本无法跟上这种秒级上下文迭代。而语义表征范式天然对齐了这一演进方向，它能将当下正在进行的对话转录文本直接作为实时上下文输入编码器，利用同一套底层语义模型无缝完成动态金融资产推荐。 在这种智能化演进过程中，设立严格的合规与风控护栏（Guardrails）成了生命线工程。在金融高风险场景中，“护栏”具有极其严苛的多重含义：最基本的要求是防止客户过度暴露在超出其风险承受能力的单一标的或特定行业板块中。当融合多模态语音交互与数字化投顾时，风险约束与意图捕捉的平衡变得更加复杂。 更关键的技术抉择在于：绝不能任由市面上的通用 LLM 在对话中自由发挥充当“金融黑嘴”。如果用户在聊天框中询问“我应该投资什么？”，后台决不能让通用模型凭空产生未经验证、充满幻觉的建议；更不能出现“对话助手推荐了一套组合，而回到 App 首页轮播图却展示着完全不同的资产”这种自相矛盾的撕裂体验。正确的架构策略是让对话式 LLM 与底层强大的金融推荐系统深度挂钩（Tool Calling / Retrieval），由统一的中央推荐引擎输出精准、受控且具备全站一致性的投资标的。为了实现这一设想，将这套高维语义推荐架构推向极端极致的“实时化”（Real-time），成了团队当前正在全力攻关的攻坚难题。 - **交互范式转变**：数字金融正从静态图形界面全面转向基于文本与语音的对话式数字投顾。 - **全站一致性保障**：对话模型必须调用底层推荐引擎，严禁 LLM 自行幻觉投资建议，确保聊天场景与 App 首页推荐完全统一。 - **合规与风控护栏**：强制嵌入口径统一的行业敞口与风险评级限制，杜绝用户资产过度暴露。 - **架构演进挑战**：实现毫秒级端到端语义推理，以满足对话交互过程中的实时推荐更新需求。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=4s) 哪只股票更好？我该买苹果还是 SpaceX？ — Alejandro Correa Bahnsen > [00:04](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=4s) 推荐不仅取决于资产本身的属性，更取决于投资者是谁。 — Alejandro Correa Bahnsen > [03:39](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=219s) 买了该资产的相似用户还买了什么 — Alejandro Correa Bahnsen > [03:39](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=219s) 接下来应该向用户展示/播放什么？ — Alejandro Correa Bahnsen > [03:39](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=219s) 用户应该如何配置/打理自己的资金？ — Alejandro Correa Bahnsen > [08:32](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=512s) 仅仅‘推荐某个资产’这一动作本身，就可能直接驱动资产价格波动。 — Alejandro Correa Bahnsen > [08:32](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=512s) 金融推荐不能仅以‘用户会点击什么’为导向，而必须建模‘该投资者在当前市场环境下最需要什么’。 — Alejandro Correa Bahnsen > [14:47](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=887s) 当特征量从 10,000 个激增到 11,000 个时，模型几乎不再吸收有效信息，反而引入了大量破坏泛化能力的噪声。 — Alejandro Correa Bahnsen > [21:48](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=1308s) 全面采用语义表征来编码用户全貌与资产画像，其潜力远远超越了过去的孤立建模路径。 — Alejandro Correa Bahnsen > [23:58](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=1438s) 绝不能任由市面上的通用 LLM 在对话中自由发挥充当‘金融黑嘴’。 — Alejandro Correa Bahnsen > [23:58](https://www.youtube.com/watch?v=Kgp27I4VNYw&t=1438s) 我应该投资什么？ — Alejandro Correa Bahnsen ## Mentioned Alejandro Correa Bahnsen, Grupo Bursátil Mexicano, GBM, Ai4, Apple, SpaceX, Google, Collaborative Filtering, Learning to Rank, LLM, Joint Representation Ranker, Recall, NDCG ## Video description Alejandro Correa Bahnsen - VP of Data & AI, Grupo Bursátil Mexicano Traditional recommendation systems such as collaborative filtering and market-signal models often fail to capture investor intent and rarely provide meaningful explanations. In this talk, I will present a financial recommender that jointly models investor behavioral signals (holdings, watchlists, interaction patterns) and market dependency structures using fine-tuned LLM-based embeddings in a shared representation space. I will explain how this joint embedding approach improves relevance, diversity, and stability compared to behavior-only or market-only baselines, and how we evaluate these gains both offline and in production-oriented settings. The talk will also cover architectural decisions, representation learning trade-offs, and explainability strategies needed for real-world deployment. If you are building ranking systems in high-stakes domains, this session will offer concrete patterns for combining foundation models with structured financial data in a scalable and interpretable way. [Source](https://www.youtube.com/watch?v=Kgp27I4VNYw)

Source quotes

[0:04] Alejandro Correa Bahnsen: 哪只股票更好？我该买苹果还是 SpaceX？ [0:04] Alejandro Correa Bahnsen: 推荐不仅取决于资产本身的属性，更取决于投资者是谁。 [3:39] Alejandro Correa Bahnsen: 买了该资产的相似用户还买了什么 [3:39] Alejandro Correa Bahnsen: 接下来应该向用户展示/播放什么？ [3:39] Alejandro Correa Bahnsen: 用户应该如何配置/打理自己的资金？ [8:32] Alejandro Correa Bahnsen: 仅仅‘推荐某个资产’这一动作本身，就可能直接驱动资产价格波动。 [8:32] Alejandro Correa Bahnsen: 金融推荐不能仅以‘用户会点击什么’为导向，而必须建模‘该投资者在当前市场环境下最需要什么’。 [14:47] Alejandro Correa Bahnsen: 当特征量从 10,000 个激增到 11,000 个时，模型几乎不再吸收有效信息，反而引入了大量破坏泛化能力的噪声。 [21:48] Alejandro Correa Bahnsen: 全面采用语义表征来编码用户全貌与资产画像，其潜力远远超越了过去的孤立建模路径。 [23:58] Alejandro Correa Bahnsen: 绝不能任由市面上的通用 LLM 在对话中自由发挥充当‘金融黑嘴’。 [23:58] Alejandro Correa Bahnsen: 我应该投资什么？

Source quotes

[0:04] Alejandro Correa Bahnsen: Which stock is better? Should I buy Apple? Should I buy SpaceX? [3:04] Alejandro Correa Bahnsen: similar customers bought item X, therefore you should buy item X [5:31] Alejandro Correa Bahnsen: What should I show next? [5:31] Alejandro Correa Bahnsen: What should I do with my money? [11:41] Alejandro Correa Bahnsen: what will the user click [11:41] Alejandro Correa Bahnsen: what is financially relevant for this specific investor right now [16:14] Alejandro Correa Bahnsen: representation gap [16:14] Alejandro Correa Bahnsen: not every interaction can be anticipated and manually created into a feature [18:35] Alejandro Correa Bahnsen: cold start problem in financial recommendations

Original transcript

[0:04] I'm here to talk you a little bit about the  evolution of financial recommendations. Um, [0:11] since I since I started working in this area  and in particular in general in asset management [0:19] and financial advisory there is I receive these  kind of questions daily. So which stock is better? [0:28] Should I buy Apple? Should I buy a SpaceX? I  mean, I get these kind of questions from friends, [0:37] from financial advisors. And the reality  is that there isn't a one answer. Yeah, [0:47] think about it. It depends in a lot of  things. It depends what why are you investing? [0:54] What are your plans? What are you looking for?  Are you just here to gamble? Are you here to put [1:01] your retirement retirement money? I mean, it has  a lot of different answers depending who you are [1:10] and of course also depending what which  is the stock that you're talking about. So [1:19] I work in a company that is almost 40 years  old traditional financial advisory company [1:27] but over the last 3 to five years we develop our  um digital app and suddenly it became very very [1:35] successful. So right now we have close  to 2 million digital users um in Mexico [1:42] is by far the large the largest brokerage uh firm  in Mexico but we keep operating as a traditional [1:51] brokerage firm um in the offline world sort of  to speak. So this is how the app looks like uh a [1:59] standard how you will see a standard investment  uh app. Uh this is how the homepage uh is and [2:10] until very recently we show the same app to  all our customers. Again we are coming from a [2:17] traditional company we move to a digital uh to a  digital application and we show exactly the same [2:24] thing to all our customers. So we have things such  as investment opportunities or popular instruments [2:32] and the way this used to be selected what to shown  to our customers. It was just every Monday morning [2:41] our chief investment officer will say this  week we are show going to show this stuff [2:48] and again this used to work when you have tens of  thousands of customers it sort of work but once [2:56] you get to millions of customers it doesn't really  work. So we started using following more or less [3:04] the uh the road map that a speaker to um the  previous speaker uh was showing us um in which [3:15] we started seeing okay what does the standard  e-commerce digital platforms do to solve this [3:24] kind of problem. We all know when we enter one of  those apps that the experience that we are seeing [3:30] is completely unique to what we are what we what  the system is is assuming that we want and so on. [3:39] So when we look at these home home screens  um it sort of look similar right? So I have [3:47] a list of items. How do I personalize which  items to show to which to which customer? [3:54] The way that you start of how you start doing  this is something that is called um collaborative [4:00] filtering. Um this is the algorithm  behind the simple algorithm behind the um [4:08] what did similar customers bought? Therefore you  should buy this. It's a very very simple algorithm [4:16] but surprisingly it works. So it only require  a little bit of math and you ended up having [4:25] what should I show to each customer. We started  doing this at the beginning of this year [4:31] and very quickly we ended up having 3% higher  sales um for the customers that we were showing [4:38] the that recommendation versus the customers that  we were showing the most popular recommendations [4:46] and and it makes sense again think about  how many different financial instruments [4:51] are out there. Um I just mentioned just  stocks but if we go more complex we have [4:57] way more instruments that you can show  to different to different customers. [5:02] So we started so this simple algorithm already  works. Um and the way that we evaluated was [5:13] how much money does the users that  were seeing this uh recommendation [5:19] ended up investing investing versus  the one that didn't that didn't saw it. [5:25] And and it's a metric that more or less makes  sense at [snorts] the beginning at least. [5:31] But but if you think about it, how should  we actually measure these kind of models? [5:37] It's not a straightforward um answer. Um when we  go to the traditional um let's say digital apps, [5:48] they will try to optimize for how much time  the person and uh is watching something. [5:55] What is the retention? What is the clicks?  uh social engagement, subscription revenue. [6:01] There there are a lot of things that are [6:05] allow me to allow me to say easy to measure and  to follow when you are recommending something [6:11] in an e-commerce or when you are recommending  something in a in in a digital music platform. [6:19] But when we are analyzing how to evaluate  recommendations of financial products, [6:27] suddenly you ended up having way more  things that you should consider. So [6:34] returns is something that  will pop up very quickly. [6:39] Of course, I want to uh give recommendations to my  customers that I have some expectations are going [6:46] to be good investments. But then you have those  customers that doesn't want to take any risk and [6:52] those customers that are here uh and and and that  are able to to take some uh some risk without or [7:02] and and will feel comfortable with that. Um so  returns it's it's a complicated one because again [7:09] no one knows what is going to be the return of  on something. The only thing that you can know [7:14] is more or less what is the volatility  of something then you want your customers [7:20] to [snorts] diversify in one way or another. If  your customer already have all the tech stocks [7:26] but should you really recommend another tech  stock? So if you are following me then you [7:32] get into the area of okay but that is what a  financial planner or a financial advisor does. [7:40] I'm not really trying to replace a  financial advisor in the home screen, [7:44] but there are a lot of things that the financial  advisor does that it should follow the same logic [7:51] in the application and and and and then you start  having realizing that the algorithms that are been [7:58] that have been developed for e-commerce  doesn't really follow that same logic. [8:06] So this the UI is the same but there  are profoundly different problems. [8:12] again in consumer optimization  what should I show next [8:17] in financial uh or in in a financial product  recommendation here the question for the user is [8:25] what should I do with my money and  it's a very very different question [8:32] there is something particularly interesting when  you are doing financial recommendations at scale [8:40] at some point. Just the fact that  you're doing the recommendation [8:45] might actually have an impact on  the price of the of the asset. [8:50] Don't think about the big um don't think about  the big uh assets. Don't think about the apples, [8:57] the Googles. Think about the niche or very small  even in a small countries kind of uh assets. [9:06] Just by showing that to a couple of million  users that previously probably have no idea [9:12] that that particular investment exist is going to  affect the price and that's very very interesting [9:20] because by all means it's not something that  we are wanting to do is just a byproduct. Um [9:27] I I try to think about in which other areas this  problem appears. Um, I haven't is is I believe [9:34] it's very very particular to financial products  recommendations. There might be others, but I [9:40] haven't get there yet. But again, this is me  [snorts] trying to show you how different it is [9:47] to show items in an e-commerce versus items um  when you are recommending people what to invest. [9:56] Another factor that affects quite a lot in this  particular problem is the temporality of your [10:01] recommendations. This of course is a problem  when you are recommending what show to show [10:08] next to a user. Of course, you don't really  want to recommend something that is years old [10:13] unless the customer is looking for  something like that, but otherwise [10:19] it's not necessarily a good recommendation. But  here what does a good recommendation means in [10:26] terms of of timing? If the pricing is just moving  a lot, do you want to show that or not? It's there [10:34] is no answer for that. So you need your model to  also being able to consider all of those things. [10:42] So so far I wanted to focus on how does  financial recommendations differ. So first [10:49] how do you measure is a completely different thing [10:52] in financial product recommendation  than on uh other e-commerce platforms. [10:59] This thing that is called price uh reflectivity  in again what I was just telling you about [11:06] how you might have an impact on the actual prices  and that will change how the recommendation looks. [11:13] And the last one that is not that difficult  uh to solve technically at least the temporal [11:19] decay. How do you consider the temporal um  the temporal aspect of your recommendations? [11:26] It is an issue in other in other  platforms here as well has um quite um [11:33] and is going to impact how do  you make these recommendations. [11:41] So moving from we need to move  from what will the users click [11:47] to what is relevant for this investor  right now under under which conditions and [12:00] and then we are we realize a recommendation is not  only talking about the asset. So you need to take [12:07] you need to understand much more and being able  to model to or to add to your model much more [12:13] information about your investor or your user. So a  useful financial recommendation is going to depend [12:20] much more in the what is the portfolio what are  the previous trades also what is the user doing [12:29] in your app. As expected, you are going to use  this kind of app to look for prices, look for [12:36] tendencies. Um, so not necessarily everything  that you ended up buying or or doing a trade, [12:43] it's neither uh is is not is going to be  relevant. You have also a lot of kind of of um [12:53] of traits that the user does and all are relevant.  You will going to have customers increasing [12:59] their position, decreasing their position. You  need to take to consider all of those things. [13:07] For some customers, there is going to be a  huge important in terms of how much taxes this [13:13] particular asset I will need to pay if I end  up having uh an important financial gain. So [13:20] even those kind of things you need to  figure out a way to add to the model. [13:27] In order to do that, our first attempt was okay,  let's model this as a machine learning problem. [13:35] So we started with a a standard algorithm, a  learning to [clears throat] rank algorithm. [13:39] And here most of the work is going to be spent  defining which are the features that you need. [13:46] Um I I I didn't put all the features here, but  you can already imagine you you have you're going [13:52] to have a lot of user features. what is going to  describe your investor. You're going to have a lot [13:58] of features regarding their current portfolio,  but then you need features from all the other [14:04] assets that that you might offer to someone  and the interaction histories with the app. [14:14] This is going to be hard work, but it's  doable. And you you're going you have [14:20] your data scientists spend several several  weeks handcrafting a lot of these features. [14:27] A lot are going to be very creative if the user  is seeing an asset that just decrease in price. [14:34] Perfect. And you will figure and you will manually  um create that kind of features. You will input [14:42] all of that into a machine learning model and  at the end you will have an output an output and [14:47] depending in which part of the app are you going  to show that um normally you will need the top [14:53] five recommendations or sort of like that in order  to personalize one of the car sales of the user. [15:01] We deployed this probably in March and after a  month or so of testing the results were very very [15:09] interesting. We ended up having almost 6% more  sales than just showing the most popular assets. [15:23] And now think about I'm doing this to millions of  users. Yeah. Is incredible the power that you can [15:31] have with this kind of of recommendations. For  sure there are a lot of things that are not yet [15:36] captured here. But even with this simple model  because again the the machine learning model [15:43] that we use is simple um most of the work was h  doing the feature engineering as we like to say [15:53] 6% increase in sales amazing um just for to  reminder to to do a reminder the first model [16:02] that was the collaborative filtering 3% here  we were able to double those those results. [16:14] We spent the last couple of months then trying to  to to keep understanding what are the limitations [16:21] of this model because results are getting there  but still there are a lot of things that we think [16:27] are missing here and there is something  that we are calling the representation gap [16:33] and that is that even though we manually create  input inside the model a lot of the signals [16:40] there a lot of other things that the  model could not really understand. [16:45] U let me give you a couple of examples. When I'm  when I'm analyzing one asset, let's say apple, [16:56] think how much information can you get from that  kind of assets. You can have a standard things [17:02] it is consumer technology. Uh what are the kind  of revenue that they are having recently? um [17:08] then if they are doing any buybacks and so on. So  a lot a lot of information that you can actually [17:16] have from an asset and the users it's similar the  kind the kind of interactions that the user have [17:25] if they what they recently bought if they  sold something as soon as the price got down [17:32] or how much did they wait when the price  was going down before selling something. [17:38] So there there are the reality that  there are really way too many features [17:44] and way too many signals that you can get  from a user. We started incorporating even [17:51] which were the news that the user was reading.  One part of our app we are showing financial news [17:58] and you are able to get that information as well.  So if the user is reading about this kind of news, [18:03] not only about this the news about an asset,  but if the news about that asset was positive [18:08] or negative, you even want to be able to  capture that kind of things in your model. [18:13] And very quickly you realize that a traditional  machine learning model is not going to be enough. [18:21] So we want to shift from doing  this financial engineering to [18:27] capturing semantic representations.  So what is a semantic representation? [18:35] We ended up creating a model that sort of looks  like this. Actually this is a very very good [18:42] visual representation of that in which we ended  up creating three kind of encoders of all the [18:50] information that we have. First of the user recent  behavior, second of the portfolio of the user [18:59] and and third for all the possible assets  that that we have available to offer someone. [19:10] After after doing that with a standard LLM,  we ended up conc concatenating all of that, [19:18] putting that into a neural network and finally  being able to extract a relevance score. [19:24] So the nice thing about modeling your problem  like this is that it's quite simple to just put [19:32] all your information as an input to an LLM, get  the encoding or the embeddings, concatenate all [19:40] of that and keep using then traditional machine  learning in order to get your recommendations. [19:49] There are many many ways to do this and and here  we are far away from deciding which is the best [19:56] way to encode this. We are just starting to work  on these kind of problems. But it was very very [20:03] interesting to see that you can leverage leverage  the power of the of how these algorithms are able [20:09] to capture knowledge and they by themselves have  a lot of knowledge of these same assets as well [20:17] and being able to put all of that together to  be a to end up having a recommendation with [20:24] way way more information about again the customer,  their portfolio and all the other assets. [20:35] A very interesting conclusion that this algorithm  allow us to have was that we were having [20:41] way too much problem recommend recommending things  that no one has purchased before. Uh think about [20:49] just a month ago we had a new huge couple of  IPOs. How do we interact with that? In the past, [20:58] we still had to manually input that into the  model. But this new kind of way [snorts] of [21:04] um creating the semantic representations  allow us to uh deal with that issue that [21:10] is normally referred as the cold start cold  start problem in financial recommendations. [21:18] So why we ended up learning the representation?  because not every interaction can be anticipated [21:25] and manually created into a feature. The  value um we keep adding features and we [21:33] got to a point in which going to from 10,000  features to 11,000 features wasn't really [21:40] adding it. It just keep adding noise. So this kind  of representation definitely help us quite a lot. [21:48] We did some experiment some experiments over  the last month and a half and ideally in the [21:55] next month we we are going to be able to put this  into production. But nevertheless, I wanted to to [22:01] to come here with you guys and show you  some offline results of this analysis. [22:08] When we measure how the the new algorithm that  we are calling a join representation ranker [22:15] compares in offline versus the other ones that I  show you before. Um measured by either by recall [22:24] or by the normalized discount community gain. um  the results are very very promising. So these kind [22:32] of representations offline at least are expecting  to have twice the result of the previous one. [22:40] And when we look by recall again how good  was the the algorithm capturing what the user [22:48] actually ended up buying, the results are quite  similar. I always prefer to come here to show [22:54] you real results but unfortunately this is  just experiments that we have been doing [22:59] and we will put this in production in  the next in the next couple of weeks. [23:06] What this is proving that the consider  that if we consider all the semantic [23:15] uh or if we encode all the information of the  user and the assets using semantic representation [23:23] potentially this is going  to have way better results. [23:27] Um again we are still in the process of doing this  on online but there are still many things that we [23:34] are not being able to understand yet and there are  more things that the actual financial advisor does [23:43] with a customer that we are in the process to  trying to figure out how we can move more of that [23:49] into a into the digital experience. But that  is going definitely to have a lot of more work. [23:58] And something that I really really like  about this kind of modeling is that [24:03] as with every financial, excuse me, as with  every digital experience, everything is moving [24:10] towards a conversational experience with your  customers, either via chat or either via voice. [24:19] The previous iteration of models  were not really suited for [24:24] tackle these kind of problems. That is as the user  is writing you something or telling you something [24:30] uh and you are responding  with a voice or chat LLM. [24:35] How do you update your recommendations? This kind  of semantic modeling actually allow you for that [24:43] because then you can input in the model in real  time the current conversation with the customer [24:48] and potentially being able to use the same model  to give the financial asset recommendation. [24:55] That is something that is just um on the  next kind of things that we are working. [25:01] Um Gabriel's is definitely an issue here and here  Gariel's means a lot of things from the purely [25:10] Gariel Gabriel's from I don't want a  customer to be overexposed to something [25:16] because his profile suggest he should not be uh  overexposed to a part to a particular sector. [25:24] How to combine that in a multimod in a in a  multimodal way particularly with boys is a huge [25:31] is very interesting and I do believe these  interactions with customers are going to be moving [25:37] way more towards uh boys with digital financial  advisors. What if the user want to want to test [25:49] convers um in a convers conversational way? what  should he invest on? So here you want to replace [25:57] the traditional LLM doing these recommendations  and have the LLM actually connecting to your [26:04] recommendation system in a way first of all  because I will not like to show a different [26:10] recommendation in a chat than in the homepage of  the app. But secondly, because the power of the [26:16] recommendations is way better to what the user  might have um in a let's say in a standard LLM. [26:25] And lastly, how do we move this towards being  more real time? It's definitely an issue, but is a [26:32] requirement for us to be able to  use it in a conversational way. [26:40] All right. Thank you so much.