AI for Investment: A Platform Disruption AI在投资中的应用：平台颠覆

Abstract

随着投资环境变得越来越竞争激烈，高效地扩大交易来源和提高交易洞察力已成为基金的主导策略。虽然基金已经在这两个任务上投入了大量努力，但它们无法通过传统方法进行扩展；因此，自动化这些任务的需求激增。最近出现了许多第三方软件提供商，他们提供生产力解决方案来满足这一需求，但它们因为缺乏针对基金的个性化、隐私限制以及软件使用案例的自然限制而失败。因此，大多数主要基金和许多小型基金已经开始开发他们自己的内部AI平台：这是行业的一个游戏规则改变者。这些平台通过与基金的直接互动变得更智能，并且可以用于提供个性化的使用案例。最近在大型语言模型方面的发展，例如ChatGPT，为其他基金开发自己的AI平台提供了机会。虽然现在没有AI平台并不构成竞争劣势，但两年后将会是。基金需要为这样的AI平台制定一个实用的计划和相应的风险评估。

With the investment landscape becoming more competitive, efficiently scaling deal sourcing and improving deal insights have become a dominant strategy for funds. While funds are already spending significant efforts on these two tasks, they cannot be scaled with traditional approaches; hence, there is a surge in automating them. Many third party software providers have emerged recently to address this need with productivity solutions, but they fail due to a lack of personalization for the fund, privacy constraints, and natural limits of software use cases. Therefore, most major funds and many smaller funds have started developing their in-house AI platforms: a game changer for the industry. These platforms grow smarter by direct interactions with the fund and can be used to provide personalized use cases. Recent developments in large language models, e.g. ChatGPT, have provided an opportunity for other funds to also develop their own AI platforms. While not having an AI platform now is not a competitive disadvantage, it will be in two years. Funds require a practical plan and corresponding risk assessments for such AI platforms.

1 Introduction: What is driving the surge for AI in the investment business? 引言：是什么推动了投资业务中对人工智能的需求激增？

随着投资环境变得越来越竞争激烈，基金正在寻求扩大其交易来源，同时提高交易洞察力，而有限合伙人（LPs）也在寻找在这一领域具有独特优势的基金。经济超级趋势导致了更多的基金和更多的待投资金，但机会和交易却更少[1]。因此，扩大交易来源并提高交易质量被认为是一种主导的获胜策略[2, 3]。通过改善他们的交易来源，基金可以更快地接触到目标资产管理团队，通过提高交易洞察力，基金可以确保更高的交易质量，并与目标管理团队进行更有针对性的对话。由于传统的交易来源和交易洞察方法成本高昂且无法扩展，基金要求使用技术，特别是人工智能，来达到这一目的。目前，基金平均花费75%的时间在交易来源和尽职调查上，这对他们来说是一个巨大的机会成本，使他们无法专注于增值任务，如建立关系和投资[3]。然而，交易来源和交易洞察无法通过传统方式进一步扩大和提高，例如与人联系、在线搜索、市场活动和打电话。因此，基金已经意识到使用自动化技术在这方面的价值[2, 4]；这还将是一个差异化的优势，在与有限合伙人交流时可以发挥作用，他们一直在寻找在这一领域具有优势的基金：75%的LPs认为人工智能对交易起源有用，64%的人相信该技术将对交易评估产生重大影响[5]。

With the investment landscape becoming more competitive, funds are looking to scale their deal sourcing while improving their deal insights, and LPs are looking for funds with a unique edge in this space. The economic megatrends have led to more funds and more dry powder but fewer opportunities and deals [1]. Consequently, scaling the deal sourcing and improving the quality of deals is considered a dominant winning strategy [2, 3]. By improving their deal sourcing, funds can get to the target asset management team faster, and by improving their deal insights, funds can ensure higher deal quality and have more focused conversations with the target management teams. Since the traditional approach for deal sourcing and deal insight are costly and fail to scale, funds are demanding technology, particularly AI, for this purpose. Currently, the funds spend on average 75% of their time on deal sourcing and due diligence, coming at a large opportunity cost for them to focus on value-added tasks such as relationship-building and investing [3]. Yet, deal sourcing and deal insights cannot be scaled and improved further in traditional ways such as connecting with people, searching online, marketing campaigns, and making calls. Hence, the funds have realized the value of using automation for this purpose [2, 4]; this will also be a differentiation that can play an advantage in talking to LPs who have always looked for funds with an edge in this space: 75% of LPs think AI is useful for deal origination, and 64% believe the technology will have significant impact on deal assessment [5].

使用替代数据来提高交易洞察力是基金对技术解决方案需求的另一个原因[6]。替代数据，例如社交媒体、评级和评论、信用卡消费、流动性、卫星图像、关键新员工招聘、新市场或产品扩张、网络流量等，是行业的一种转变，特别是在私募股权和风险资本领域，需要使用人工智能[7]。虽然对冲基金和公共资产投资者已经认识到并整合了替代数据在产生alpha（超额回报）方面的价值，私募股权和风险资本公司最近才开始尝试使用它们。然而，这些替代数据源通常是庞大、非结构化和复杂的，这使得人类无法有效分析并产生洞察力。幸运的是，机器学习算法可以分析这些数据，进行模式匹配，并从噪声中找到信号[6]。

The use of alternative data for improving deal insights is another reason for funds’ demand for technology solutions [6]. Alternative data, for example, social media, rating and review, credit spending, mobility, satellite image, key new hires, new market or product expansion, web traffic, etc., is a shift in the industry, especially in private equity and venture capital, and requires the use of artificial intelligence [7]. While hedge funds and public asset investors already appreciate and incorporate the value of alternative data for generating alpha, private equities, and VCs have more recently started to experiment with them. These alternative data sources however are often massive, unstructured, and complex, which makes them impossible for a human to effectively analyze and generate insights. Fortunately, machine learning algorithms can analyze this data, pattern match, and find the signal from the noise [6].

2 Why third party software solutions fail at addressing the AI demand in investment business? 为什么第三方软件解决方案无法满足投资业务中对人工智能的需求？

由于上述动态，近年来出现了第三方解决方案提供商，以满足基金使用技术进行投资的需求。这些第三方解决方案是为基金提供的数据生产力工具，形式上是软件。在交易来源和交易洞察的领域内，这些解决方案可以分为三类：a) 聚合公司数据库及其主要原型，加上在数据库中进行搜索的能力。b) 聚合和处理替代数据以获得交易洞察。c) 通过组织流程并提供对问题的现成回答来提供标准的尽职调查。虽然这些第三方解决方案可以提高生产力，但它们未能满足基金的实际需求[8]。这主要有四个原因：缺乏个性化、智能有限、隐私问题和软件的自然限制。

Consequent to the above dynamics, there has been an emergence of third party solution providers in recent years to address the funds’ need for using technology for investment. These third party solutions are in the form of data productivity tools for the funds that are offered as software. Within the space of deal sourcing and deal insight, these solutions can be categorized into three groups: a) aggregating a database of companies with their major archetypes, plus a search capability across the database. b) aggregating and processing alternative data for deal insights and c) providing standard due diligence by organizing the process and providing ready responses to questions. While these third-party solutions can boost productivity, they fail to address the real needs of the funds [8]. This is the case for four primary reasons: Lack of personalization, limited intelligence, privacy concerns, and natural limits of software.

• 缺乏个性化：现有的解决方案没有针对基金进行个性化定制，因此需要大量的人工工作来为基金创造附加价值。换句话说，这些解决方案不了解基金的投资理念及其独特的优势和需求。相反，它们的结果是通用的。然而，实际上，每个基金在其投资理念、方法和投资组合上都是独一无二的，理解这种独特性对于驱动价值至关重要。将这些机器的结果转化为基金的实际价值需要大量的人力努力。

• 智能有限：由于无法访问基金数据，也导致了智能和学习能力的有限。对于一个机器学习算法，例如推荐算法，要达到竞争性能，它应该有一个活跃的数据管道，从基金中获取数据以变得更智能。例如，它应该从基金接收反馈，导航它缺少哪些数据以主动收集它们，甚至使用投资组合公司中可用的数据。由于第三方软件解决方案无法获得这些，它们无法随着时间的推移变得更智能。

• 隐私问题：将基金数据提供给第三方带来了巨大的隐私风险。这种风险的一个例子是将根据保密协议收集的目标资产数据上传到第三方数据库。实际上，这种隐私问题也是第三方解决方案无法个性化的原因之一。

• 软件的自然限制：每个软件解决方案只覆盖了这个碎片化空间的一部分，并且是以不完整的方式。软件天然是为某些用例设计的，不能扩展到基金中的其他用例。这意味着为了解决其多样化和不断出现的自动化需求，基金将不得不获得一系列软件的访问权限，并尝试将它们全部整合在一起。这将导致繁琐的流程和低效的结果。

• Lack of personalization: The existing solutions are not personalized to the fund and hence require major manual work to generate added value for the funds. In other words, these solutions do not know the fund investment thesis and its unique strengths and requirements. Rather their outcome is generic. However, in reality, every fund is unique in its investment thesis, approach, and portfolio, and understanding such uniqueness is critical for driving value. It takes a large human effort to turn the outcome of these machines into real value for the fund.

• Limited intelligence: Lack of access to the fund data also results in limited intelligence and ability to learn. For a machine learning algorithm, e.g., a recommendation algorithm, to achieve competitive performance, it should have an active pipeline of data from the fund to grow more intelligent. For example, it should receive feedback from the fund, navigate which data it is missing to actively collect it, and even use the data available at the portfolio companies. Since none of these is available to third party software solutions, they fail to get more intelligent over time.

• Privacy concerns: Providing the fund data to a third party comes at a large privacy risk. An example of such risks is uploading data gathered under NDA from a target asset to a third party database. In fact, this privacy concern is one reason why third party solutions can not become personalized.

• Natural limits of software: Each software solution only covers part of this fragmented space and in an incomplete way. Software naturally is designed for certain use cases and cannot be extended to other use cases in the fund. This means that to address its diverse and emerging needs for automation, the fund will have to get access to a collection of software and try to put them all together. This will lead to cumbersome processes and inefficient outcomes.

由于第三方软件解决方案在生产力工具方面的限制，行业已经出现了向数据驱动平台解决方案的转变。

Due to the limits of third party software solutions for productivity tools, there has been a shift in the industry toward a data-driven platform solution.

3 What is the AI platform disruption in the investment business?

投资业务中的人工智能平台颠覆是什么？

由于第三方解决方案存在不足，几乎所有主要基金和一些小型基金都开始构建内部专有的自动化平台，用于他们的交易来源和交易洞察：这是行业中的一个游戏规则改变者[8, 9, 10, 11]。一些值得注意的例子包括EQT Motherbrain机器、黑石数据科学团队（BXDS）、贝莱德Aladdin机器、Advent实验室和KKR数据科学团队。这些机器主要参与推荐新交易和收集交易洞察，并且已经引领了许多成功的交易。还有超过80家风险资本公司开始构建内部自动化机器，预计到2025年这一比例将增长至75%的风险资本基金[2]。其他类型的投资基金也存在类似的趋势。这些自动化机器/人工智能平台与传统软件解决方案在三个重要方面有所不同。

Consequent to the shortcomings of the third party solutions, almost all major funds and some smaller funds have started building in-house proprietary automation platforms for their deal sourcing and deal insights: a game changer in the industry [8, 9, 10, 11]. Some notable examples are the EQT Motherbrain machine, Blackstone data science team (BXDS), Blackrock Aladdin machine, Advent lab, and KKR data science team. These machines are primarily involved in recommending new deals and gathering deal insights and have already led many successful deals. There are also more than 80 VCs that have started building in-house automation machines which is expected to grow to 75% of the VC funds by 2025 [2]. Similar trends exist in other types of investment funds. These automation machines / AI platforms are different from traditional software solutions, in three important respects.

• 数据集成平台：这些机器作为一个平台，收集并连接所有内部数据和外部数据到一个智能系统中。机器可以访问基金的专有数据，例如投资会议或之前的尽职调查，通过这种方式了解基金独特的投资方法。重要的是，这些机器随着时间的推移会收集越来越多关于基金的信息，从而变得更加智能。在这个意义上，投资者与人工智能机器的每一次互动都会使它们变得更聪明。此外，访问来自投资组合公司的选择性数据，例如客户偏好、销售、供应链等，可以帮助机器深入洞察其相应的行业和趋势。

• 个性化平台：随着机器对基金的了解，其结果会针对基金进行个性化。例如，为基金推荐的交易会考虑基金的投资理念、其在从资产中创造价值方面的独特优势和优势，以及投资组合的当前状态。同样，交易洞察会通过考虑基金的独特优势来识别目标资产的独特风险和上行潜力。

• 多样化用例平台：基金可以使用这个平台以一种连贯的方式自动化不同的用例。这样，基金不需要多个软件，而是有一台可以随着时间推移来解决不同用例的机器。随着人工智能技术的发展，这一特性尤为重要，它可以使不同基金功能的自动化变得可行。此外，每个功能都可以有多个用例，这些用例都可以集成到这一内部机器中。

• Data integration platform: These machines act as a platform to collect and connect all internal data and external data in one intelligent system. The machine has access to the fund’s proprietary data, e.g. investment meetings or previous due diligence, and this way learns about the fund’s unique approach to investment. What is important is that these machines gather more and more information about the fund over time, and this way become more intelligent about it. In this sense, every interaction of the investors with the AI machine makes them smarter. Furthermore, access to the selected data from portfolio companies, e.g. customer preference, sales, supply chain, etc., can help the machine to generate deep insights into their corresponding industry and trends.

• Personalization platform: As the machine learns about the fund, its outcomes are personalized to the fund. For example, the recommended deals for the fund are by considering the fund investment thesis, its unique strength and advantages in generating value from an asset, and the current state of the portfolio. Similarly, the deal insights identify the unique risks and upsides of a target asset by considering the fund’s unique advantage.

• Platform for diverse use cases: The fund can use this platform to automate the different use cases all in a coherent way. This way, instead of multiple software, the fund has one machine that can grow over time to address different use cases. This is a particularly important feature with growing AI technology which can make automation of different fund functions feasible over time. Furthermore, each function can have multiple use cases which can be all integrated into this one internal machine.

在理解这些平台所驱动的价值时，首先应该注意到，这些平台是对人类性能的增强，旨在提升人类的表现。其次，应该考虑到这个平台是用于整个投资阶段的真正数据驱动用例。

In understanding what the value driven by these platforms is, one should first note that the platforms are an augmentation to boost human performance. Second, one should consider that the platform is for true data-driven use cases across all stages of the investment.

• 首先，投资领域的人工智能工具是对人类性能的增强。换句话说，人类将与机器合作，形成混合智能，以做出投资决策。实际上，研究表明，一个配备了机器的普通人类可以胜过行业内经验丰富的明星[12]。通过提升人类性能，从分析师到合伙人，机器使他们能够将时间花在真正的增值任务上，即建立关系和进行投资。

• 其次，内部专有的人工智能机器为基金提供了一个平台，使其能够从整个投资周期中的不同用例中提取价值并受益。目前，这些机器主要用于交易来源，包括并购选项，以及与之密切相关的交易洞察：任何推荐都应伴随着清晰的洞察和推理。与第三方软件解决方案不同，这些平台的用例不仅限于生产力工具，也不仅限于投资周期的开始阶段，例如搜索公司数据库或使用内部训练的ChatGPT进行标准尽职调查。以下是整个投资周期中用于交易来源的真正数据驱动用例的一些例子：

• First, the AI tool for investment is an augmentation to boost human performance. In other words, humans will work with the machine to form hybrid intelligence in making the investment decisions. In fact, research has shown that a descent human augmented with a machine can outperform seasoned stars of the industry [12]. By boosting human performance, from analysts to partners, the machine allows them to spend time on real value-added tasks i.e. building relationships and making investments.

• Second, the in-house proprietary AI machine provides a platform for the fund to derive and benefit from different use cases across the investment cycle. Currently, these machines are primarily used for deal sourcing, including M&A options, and deal insights which are hand-in-hand: any recommendation should be followed by clear insights and reasoning. As opposed to the third party software solutions, the use cases from these platforms are not limited to productivity tools and only at the beginning of the investment cycle e.g. database of companies to search or doing standard due diligence with a ChatGPT trained in-house. Some examples of true data-driven use cases for deal sourcing across the investment cycle are the following:

– 机器通过利用主流投资数据（例如Crunchbase、Pitchbook、Sourcescrub等）、替代数据（例如大量客户评论、情绪、流动性、信用卡等）以及基金或其投资组合公司中的任何专有数据，为基金推荐专有交易或并购机会。推荐可以直接来自投资者的原型，或者更有趣的是，通过自由探索当前投资空间的邻近子行业以及具有可转移知识/专长的行业空间。

– 机器识别选定资产的子行业，并找到所有相似或相近的公司作为替代选项。

– 机器为推荐的交易提供个性化的评分、风险和上行潜力。

– 机器跟踪资产以推荐购买时机。它识别领先卖出指标以决定何时购买选定的资产，例如新的C级管理层、新的筹资活动、组织变动等，生成频繁的报告，并使用领先卖出指标在正确的时间发送购买提醒。

– The machine recommends proprietary deals or M&A for the fund by leveraging mainstream investment data, e.g. Crunchbase, Pitchbook, Sourcescrub, etc, alternative data e.g. large customer review, sentiment, mobility, credit card, etc., and any proprietary data at the fund or its portcos. The recommendation can be by taking archetypes directly from investors, or more interestingly by freely exploring the space including adjacent sub-sectors to current investment space, and industries with transferable knowledge/expertise.

– The machine identifies the sub-sector of a selected asset and finds all similar or closeby companies as alternatives.

– The machine provides personalized scores, risks, and upsides about the recommended deals.

– The machine tracks the asset for recommending a buy time. It identifies lead sell indicators for making a move on buying a selected asset, e.g. new C-level, new fundraising, organizational changes, etc., generates frequent reports, and sends an alert on the correct time to buy using the lead sell indicators.

同样地，整个投资周期中用于交易洞察的真正数据驱动用例的一些例子包括：

Similarly, some examples of data-driven use cases for deal insight across the investment cycle are the following:

– 机器通过考虑基金的能力、投资组合和投资理念，并使用主流、替代和专有数据，识别出在评估目标时需要关注的个性化关键特征和警示信号。

– 机器采用多方面的、健全的方法提供关于评估目标的关键特征和关键风险的详细洞察，例如使用结构化方法，即文档处理，以及通过解释推荐系统中使用的神经网络的特征来采用非结构化方法。

– 机器通过提出一系列数据请求和来自目标的问题，为管理层演示会议构建结构，以便在流程的早期进行有针对性的对话并更快地获得智慧。

– 机器接收目标提供的多样化私有数据，并更新管理层演示中的相应关键特征、风险因素、洞察和下一步行动。

– 机器进一步接收目标提供的多样化私有数据，并更新管理层演示中的相应关键特征、风险因素、洞察和下一步行动。

– 机器在投资委员会会议上为投资某资产提供一个非二元的评分/投票。

– The machine identifies personalized key features and red flags to focus on in assessing a target. These are by considering the fund capabilities, its portfolio, and investment thesis, and using mainstream, alternative, and proprietary data.

– The machine provides detailed insights on key features and key risks for assessing a target using a multi-faceted and robust approach e.g. using a structured approach i.e. document processing, as well as an unstructured approach by interpreting the features of the neural network used in the recommendation system. – The machine structures the management presentation meetings for focused conversations and getting smart earlier in the process, by proposing a list of data requests and questions from the target.

– The machine intakes diverse private data provided by the target and updates the corresponding key features, risk factors, insights, and next steps with the management presentations

– The machine further intakes diverse private data provided by the target and updates the corresponding key features, risk factors, insights, and next steps with the management presentations.

– The machine provides a non-binary score/vote in the investment committee meeting for investment in an asset.

上述用例表明，人工智能机器从投资的初始步骤到最终决策阶段都为投资者提供支持。

The above example use cases show that the AI machine provides support to the investors from the beginning step of the investment to the final decision-making.

值得注意的是，这些内部AI平台的激增得到了投资流程中人工智能技术进步的支持[13]。有超过200篇近期的学术论文，其中一些直接由投资基金支持，提供了自动化投资流程的算法（有关文献综述见[14]）。这些论文提出了并成功测试了机器学习算法，使用神经网络来预测公司的成功、为特定基金推荐投资、从替代数据中生成洞察以及对投资进行推理。这一领域的快速进展导致了更复杂的算法，这些算法能够模拟整个行业生态系统，并识别基金投资理念在这一空间中的位置（例如见[15, 16, 17]）。这些算法的成功为使用它们的基金提供了足够的概念证明，确保了它们的实际价值。

It is important to note that the surge of these in-house AI platforms has been backed by AI technological progress for investment processes [13]. There are more than 200 recent academic research papers, some directly supported by investment funds, that provide algorithms for the automation of the investment processes (for a literature survey see [14]). These papers have proposed and successfully tested machine-learning algorithms, using neural networks, in predicting the success of a company, recommending investments for a specific fund, generating insights from alternative data, and reasoning about an investment. The rapid progress in this space has led to more complex algorithms that are able to model the entire industry ecosystem and recognize where in this space the fund investment thesis lies (e.g. see [15, 16, 17]). The success of these algorithms has provided sufficient proof of concept to ensure real value for the funds using them.

4 How have large language models changed the game for AI platforms in the investment business?

大型语言模型如何改变了投资业务中AI平台的游戏规则？

ChatGPT、PaLM、Falcon、Cohere以及其他类似的开源大型语言模型加速了投资领域AI平台的颠覆性变革，并允许中小规模基金也加入到开发内部专有平台以重新获得其竞争优势的游戏中。

尽管普遍的看法认为，这些模型的真实影响远远超出了仅仅使用ChatGPT作为一个聊天机器人来询问有关资产的标准尽职调查问题。实际上，这些模型可以在AI平台的不同部分使用，以显著提高其性能（例如[18]）。因此，这些开源大型语言模型的影响可以分为以下6个领域。

ChatGPT, PaLM, Falcon, Cohere, and other similar open-source large language models have fastforwarded the AI platform disruption in investment, and allow small and mid-size funds to also enter the game for recapturing their competitive advantage by developing in-house proprietary platforms. Despite the common perception, the real impact of these models is far beyond just using ChatGPT as a chatbot for asking standard due diligence questions regarding an asset. In fact, these models can be used in different parts of the AI platform to significantly improve its performance (for example [18]). As such, the impact of these open-source large language models can be categorized into the following 6 areas.

• 数据准备/清洗：这些大型语言模型通过其结构化非结构化数据的能力显著降低了数据清洗的成本。许多基金没有一个有组织且结构良好的数据库供其平台使用，而为机器学习算法准备数据的过程既耗时又昂贵。大型语言模型可以接收多样化的数据并将它们放入预定义的结构中。

• 模型训练：这些模型减少了为投资用例训练机器学习模型所需的成本和时间。训练和调整机器学习算法以达到每个用例的高性能是一项复杂的任务。然而，大型语言模型提供了一个基础的巨型模型，可以更快、更经济地重新训练以适应每个特定的用例。

• 专有数据个性化：这些模型减少了驱动基金独特价值所需的专有数据量。虽然从头开始训练一个神经网络模型以理解基金的指纹需要大量的专有数据，但大型语言模型可以重新训练，即使用显著更少的数据提供更好的洞察。

• 隐私保护：这些模型提供了隐私保护解决方案。虽然第三方解决方案总是存在数据泄露的风险，但大型语言模型的开源性质允许对其进行分支化，并在基金服务器上使用它们。这样，模型及其所有相关训练数据都安全地保存在基金内部。

• 文本嵌入：这些模型可以显著提高文本嵌入的质量，这是投资流程中机器学习算法使用的一个主要功能。文本嵌入允许将关于资产或基金的文本输入转换为数值向量，然后由机器使用。这种嵌入的质量是机器学习算法质量的决定性因素。大型语言模型特别擅长处理文本并生成此类嵌入。

• 解释神经网络：这些模型可以用来解释神经网络的特征，这是生成关于推荐交易洞察所需的功能。神经网络是AI引擎的主要工具，能够识别数据中的模式。然而，使用神经网络的一个主要挑战是它们的不可解释性。换句话说，神经网络机器选择的特征不易解释。而在投资业务中，任何推荐都应该附有清晰的理由和洞察[19]。大型语言模型为解释神经网络的特征提供了一种新方法，可以用于推理推荐的交易并提供洞察。

• Data preparation/cleaning: These large language models have significantly reduced the cost of data cleaning with their ability to structure unstructured data. Many funds do not have an organized and well-structured database to use for their platform, and the process of preparing data to be used for the machine learning algorithm is time-consuming and costly. Large language models can intake diverse data and put them into predefined structures.

• Model training: These models have reduced the cost and time required for training machine learning models for the investment use cases. Training and tuning a machine learning algorithm to achieve high performance for each use case is a complicated task. The large language models however have provided a base mega model that can be retrained for each specific use case in a faster and more cost-efficient way.

• Proprietary data for personalization: These models have reduced the amount of proprietary data required for driving value uniquely to the fund. While training a neural network model from scratch in a way to understand the fund’s fingerprint requires a large amount of proprietary data, large language models can be retrained to provide even better insights with significantly smaller data.

• Privacy preservation: These models have provided privacy-preserving solutions. While third-party solutions are always at risk of data leakage, the open-source nature of the large language models allows forking them and using them in the fund servers. This way, the model and all the data associated with training it are kept at securely inside the fund.

• Text embedding: These models can significantly improve the quality of text embedding, a major function used in machine learning algorithms for investment processes. Text embedding allows turning text inputs about assets or funds into a numerical vector which can be then used by the machine. The quality of such embedding is a determining factor for the quality of the machine learning algorithm. Large language models are particularly good at processing text and generating such embedding.

• Interpreting neural networks: These models can be used to interpret the features of a neural network, a function that is required for generating insights regarding the recommended deals. Neural networks, the main tool in AI engines, are black boxes that can recognize patterns in data. However, a major challenge in using neural networks is their lack of interpretability. In other words, the features selected by the neural network machine are not easy to interpret. This is while in the investment business, any recommendation should be supplemented with clear reasons and insights [19]. Large language models have provided a new approach to interpreting the features of neural networks, which can be used for reasoning about the recommended deals and providing insights.

虽然大型语言模型在投资行业的早期测试阶段，但它们已经对AI平台显示出了重大的改进，并承诺在未来带来更多的改进。

While the large language models are still in their early test phase for the investment industry, they have already shown major improvements to the AI platforms and are promising more improvements in the future.

5 Where should a fund start practically for an AI platform?

基金应该如何实际着手开发一个AI平台？

对于开发内部AI平台的实用方法，基金应该回答以下六个问题：

• 基金特定的AI平台需要什么要求/功能？

• 使用平台进行自动化的优先用例是什么？

• 哪些数据被输入到AI平台中？

• 平台中使用了哪些模型和算法？

• 谁是开发团队，即是自建还是购买？

• 开发的成本是多少？

For a practical approach to developing an in-house AI platform, a fund should answer six questions:

• What are the requirements from/features of its AI platform specific to the fund?

• What are the priority use cases for automation using the platform?

• What are the data fed into the AI platform?

• What are the models and algorithms used in the platform?

• Who is the development team i.e. build-vs-buy?

• What is the cost of development?

下面我们将详细讨论这六个问题。

We elaborate on each of these six questions separately below.

AI 平台的特性： 基金首先应该定义其平台所需的一系列特性，以为其开发奠定基础。虽然每个基金的AI平台可能具有独特的特性，但成功的AI平台应该具备一些基本特性。这些特性包括自适应学习、个性化、隐私保护和向前兼容性。自适应学习确保机器有一个活跃的数据收集管道，无论是内部还是公共数据，以增强其智能。这包括从基金中的投资者那里接收积极的反馈。个性化确保在所有用例中，AI机器都能提供独特于该基金并与其投资理念相匹配的结果，而不是通用结果。隐私保护是确保机器在使用每一块私有数据时都有明确的界限，并确保数据不会外泄，也不会在内部不同实体之间泄露，例如在不同保密协议下共享数据的不同资产。最后，向前兼容性确保机器技术能够采用AI行业的最新发展，例如推荐算法的新进展或新的开源基础模型。

Features of the AI platform: A fund should first define the set of features required from its platform to set the foundation of its development. While each fund’s AI platform may have unique features, there are a few essential features that a successful AI platform should have. These features include adaptive learning, personalization, privacy-preserving, and forward compatibility. Adaptive learning ensures that the machine has an active pipeline of gathering valuable data, either internal or public, to enhance its intelligence. This includes receiving active feedback from the investors in the fund. Personalization ensures that in all use cases, the AI machine is delivering outcomes unique to the fund and matched to its investment thesis rather than generic outcomes. Privacy-preserving is to ensure the machine has clear boundaries on how it uses each piece of private data and ensures no leakage of the data either externally, or internally between different entities e.g. different assets sharing data under different NDAs. Finally, forward compatibility ensures the machine technology can adopt the latest developments in the AI industry, such as new advancements in the recommendation algorithms or new open-source foundation models.

优先用例： 其次，基金应从其AI平台中选择优先用例。在选择这些用例时有两个经验法则。

• 基金应确保捕捉到目前可用的低风险高回报结果，这些结果是通过最先进的技术实现的。交易来源和交易洞察，两者都针对基金进行个性化，在这方面是优先考虑的。如前所述，目前的投资推荐系统技术和研究已经为这两个用例提供了足够的概念证明。此外，ChatGPT和其他开源大型语言模型可以显著提高这些推荐系统的质量。基金正确定义交易来源和交易洞察的相应用例，以从其平台中获取AI的真正价值，这一点非常重要。

• AI平台用例不应仅限于早期资产发现，而应继续在整个投资决策过程中提供洞察和支持投资者。实际上，随着决策过程接近最终投资决策阶段，AI机器在提供目标资产的补充洞察方面更有价值。

Priority use cases: Second, the fund should select the priority use cases from its AI platform. There are two rules of thumb in choosing those use cases.

• The fund should ensure it captures those low-risk high-return outcomes that are currently available with state-of-the-art technology. Deal sourcing and deal insight, both personalized to the fund, are priorities in this sense. As mentioned before, the current state of the technology and research in recommendation systems for investment have provided sufficient proof of concept for these two use cases. Furthermore, ChatGPT and other opensource large language models can improve the quality of these recommendation systems significantly. It is important that the fund correctly defines the corresponding use cases for deal sourcing and deal insight to garner the true value of AI from its platform.

• The AI platform use cases should not be limited to early-stage asset discovery, and rather they should continue providing insights and supporting the investors through the process to the last stage of the investment decision. In fact, the AI machine is even more valuable in providing complementary insight to the target asset as the process gets closer to the final investment decision.

投入人工智能平台的数据： 基金应确保机器从多样化的来源持续获取数据，以实现竞争优势。通常，数据来源包括三种类型：a) 公共数据，包括标准数据，如Crunchbase、Pitchbook等，以及替代数据，如评论、社交媒体、信用支出等；b) 专有数据，包括内部数据和投资组合公司数据；c) 根据保密协议从目标资产提供的私有数据。对于这三种来源，基金应仔细调查哪里存在有价值的数据，并建立数据收集管道来收集这些数据。重要的是，机器不断从每个来源接收更多数据，以增强其智能并更明智地了解投资生态系统。此外，应不断监控机器在数据不足需要增强的领域。这样，机器才能实现最高的竞争优势。

Data fed into the AI platform: The fund should ensure the machine has a continuous pipeline of data from diverse sources to achieve a competitive edge. Generally, the data sources include three types: a) public data including standard data, e.g. Crunchbase, Pitchbook, etc. and alternative data, e.g. reviews, social media, credit spent, etc., b) proprietary data including internal data and portfolio company data, and c) private data from its target assets provided under NDA. For each of these three sources, the fund should carefully investigate where the valuable data exists, and put in place a datagathering pipeline to collect those data. It is important that the machine continuously receives more data from each source to enhance its intelligence and grow smarter about the investment ecosystem. Moreover, the machine should be constantly monitored for areas with a lack of sufficient data to be enhanced. This way the machine can achieve the highest competitive advantage.

模型和算法： 在选择开发人工智能平台的工程架构和相应的模型/算法时，基金应确保使用多样化的算法，并通过开源大型语言模型（如ChatGPT）的真实力量来增强它们。投资用例，如交易来源和交易洞察，本质上是多方面的任务，应以强大的方式完成；没有单一的算法总是最好的。这样，基金应确保使用多种多样化的算法来产生最佳结果。其次，如上所述，如果正确使用并充分发挥其全部能力，开源基础模型对这些算法的性能具有变革性的影响。

Models and algorithms: In choosing an engineering architecture and the corresponding models/algorithms for developing the AI platform, the fund should ensure the use of diverse algorithms and enhance them with the true power of open-source large language models e.g. ChatGPT. The investment use cases, such as deal sourcing and deal insight, are by nature multi-faceted tasks that should be done in a robust way; no single algorithm can always be the best. In this way, the fund should ensure the use of multiple diverse algorithms to generate the best outcome. Second, as mentioned above, the open-source foundation models are transformative to the performance of these algorithms if used correctly and at their full capacity.

开发团队，即自建与购买： 在选择构建人工智能平台的团队时，基金应考虑三种能力：技术能力、投资流程的专业知识和构建能力。技术能力不仅包括工程技能，还包括高级研究。用于投资用例的最先进人工智能架构和算法随着许多学术出版物的发展而增长。技术团队应能够与这些研究建立联系。投资流程的专业知识是定义增加投资价值的正确和明确的用例所必需的。最后，构建这样一个平台需要与投资基金的利益相关者紧密合作。这包括定义用例，通过与机器互动收集反馈，开发用户界面，以及培训他们使用机器。以顺畅无摩擦的方式处理这些互动，对基金的日常运作至关重要，这是构建成功的关键。基金可以为此目的建立一个内部团队，或者外包构建。虽然内部构建可以导致在内部建立工程能力，但外包解决方案可以更快、成本更低、对基金的摩擦更小。外包的另一个重要好处，特别是对于管理资产不到500亿美元的基金，是基础平台基础设施可以作为公共事业运营，所有参与者的使用带来的学习和智能收益都归所有最终用户。这种方法最大化了学习的速率和机器的整体智能，同时谨慎地实现了基金的隐私和差异化。下一段将对此进行更详细的描述。

Development team i.e. build-vs-buy: In selecting a team for building the AI platform, the fund should consider three capabilities: technical capability, know-how of investment processes, and ability to build along. The technical capability includes not only engineering skills but also advanced research. State-of-the-art AI architectures and algorithms for investment use cases are growing with many academic publications. A technical team should be able to connect to this body of research. The know-how of the investment processes is required for defining the correct and sharp use cases for adding value for investment. Finally, building such a platform requires working closely with the stakeholders in the investment fund. This includes defining the use cases, gathering feedback by interacting with the machine, developing the user interface, and training them to use the machine. Handling these interactions in a smooth and frictionless way to the day-to-day work of the fund is key to a successful build. A fund can either build a team in-house for this purpose or outsource the build. While building in-house can result in building the engineering capability internally, the outsourcing solution can be faster, at a lower cost, and with less friction for the fund. Another important benefit of outsourcing, especially for funds less than $50B asset under management, is that the base platform infrastructure can operate as a public utility, with learnings and intelligence gains from all participant usage accruing to all end users. This approach maximizes the velocity of learning and overall intelligence of the machine, while carefully enabling fund privacy and differentiation. More on this will be described in the next paragraph.

开发成本： 虽然这样的定制内部人工智能平台的成本与SaaS解决方案不可比，考虑到大型基金为其预算数千万美元，但有两个因素可以为新加入者降低成本。首先，如上所述，如ChatGPT这样的开源基础模型可以减少构建平台的时间和成本。其次，工程解决方案由三层组成，其中两层是基金独立的，可以在基金之间共享。这三层是基础设施层、公共数据训练层和个性化层。基础设施层包括准备神经网络算法用于推荐、行业图建模、数据收集管道、与ChatGPT和其他大型语言模型合作的API等。公共数据训练层包括用包括标准数据和替代数据在内的投资知识训练模型。个性化层是价值的主要驱动力，包括在基金安装基础设施、构建其专有数据、定义其平台特性和用例、为其用例训练算法、培训员工，并将解决方案嵌入现有的IT解决方案中。虽然前两层是基金独立的，可以在基金之间共享，但个性化层是为基金量身定制的，增加了主要价值。

Cost of development: While the cost of such an in-house bespoke AI platform is not comparable with SaaS solutions, considering larger funds are budgeting tens of millions of dollars for it, there are two factors that can reduce the costs for the new joiners. First, as mentioned above, the open source foundation models such as ChatGPT can reduce the time and cost of building a platform. Second, the engineering solution consists of three layers, two of which are fund-independent and can be shared across funds. These three layers are the infrastructure layer, the public data training, and the personalization layer. The infrastructure layer includes preparing the neural network algorithms for recommendation, graph modeling of the industry, data gathering pipeline, APIs for working with ChatGPT and other large language models, etc. The public data training layer includes training the models with investment knowledge including standard data and alternative data. The personalization layer is the main driver of value and includes installing the infrastructure at the fund, structuring its proprietary data, defining its platform features and use cases, training algorithms for its use cases, training employees, and embedding the solutions into the existing IT solutions. While the first two layers are fund-independent and can be shared across funds, the personalization layer is bespoke to the fund and adds the main value.

6 How the risks can be evaluated in taking any AI response?

如何评估采用任何AI回应所面临的风险？

采用AI回应的最大风险实际上是不采取任何行动。行业转变已经开始，大多数主要基金已经开始开发内部工具，许多较小的基金也在进行类似的开发。这项技术已经度过了其S曲线的发酵阶段，开始进入起飞阶段。两年内，许多基金会拥有这样的AI平台。因此，虽然目前一个基金没有这样的AI平台并不是竞争劣势，但两年后就会是[13]。

另一方面，早期行动者将保持领先地位。与其他技术领域早期行动者面临探索风险不同，在AI领域，早期行动者因为解决方案的个性化特性而拥有优势。一个基金为其AI回应准备基础设施、清理历史数据、学习其独特的使用AI的方式，并在使用技术方面进行文化转型需要时间。虽然大型语言模型使得新基金更容易进入这一领域，但两年后可能已经太晚，无法进入竞争以重新获得竞争优势。

在使用AI工具时，一个主要的风险类别是隐私和合规性。如上所述，AI平台应具备保护隐私的功能。为此有三种标准做法：首先，基金在训练专有和私有数据时应开发其内部机器。内部机器安全地保存在基金服务器上，以防止数据的外部泄露。其次，基金在处理不同资产下不同NDA（保密协议）的数据时，应训练具有明确界限的独立模型。NDA时间线终止后，这些模型应相应地被删除。这种做法禁止使用一个资产的数据来生成另一个资产的洞察。最后，对于高度敏感的数据，例如人力资源数据，基金可能会考虑不直接在这些数据上训练模型，而是仅仅使用机器在这些文件中搜索所需的结果。这样，每个资产的敏感数据在其自身的洞察中如何使用就变得清晰了。虽然这三点是基本做法，但基金还应考虑更多的隐私和合规性做法。

The largest risk in taking an AI response is not taking any action. The industry shift has already started, most major funds have started developing in-house tools and many smaller funds are doing similar developments. The technology has passed the fermentation phase of its S-curve and is starting its take-off phase. In two years many funds will have such AI platforms in-house. Hence, while it is not a competitive disadvantage for a fund to not have such an AI platform now, it will be one in two years [13].

On the other hand, the early movers will stay ahead of the curve in this space. Unlike other technology sections where early movers face an exploration risk, in this AI space, early movers have an advantage because of the personalized nature of the solutions. It takes time for a fund to prepare the infrastructure for its AI response, clean its historical data, learn its unique way of using AI, and have the cultural transformation in using the technology. While the large language models have made it far more accessible for new funds to enter this game, in two years it may be too late to enter the competition to recapture the competitive advantage.

One major category of risk in using AI tools is privacy and compliance. As mentioned above, an AI platform should have the privacy-preserving feature. There are three standard practices for this purpose: First, the fund should develop its in-house machine whenever training proprietary and private data. The in-house machine is kept on the fund servers securely to stop external leakage of the data. Second, the fund should train separate models with clear boundaries when working with the data from different assets under different NDAs. The models should consequently be deleted after the NDA timeline terminates. This practice forbids using data from one asset to generate insights on another asset. Finally, for highly sensitive data, e.g. HR data, the fund may consider not training the model directly on those and rather just use the machine to search through those documents for required outcomes. This way it is clear how the sensitive data of each asset is used within its own insights. While these three are the basic practices, there are more privacy and compliance practices that a fund should consider.

7 Conclusion: What is the future of AI in investment?

结论：AI在投资领域的未来是什么？

AI平台在投资领域的颠覆已经开始，目前正处于S曲线的发酵阶段末期，以及起飞阶段的初期。大多数主要基金和许多小型基金已经开始开发这样的内部专有AI平台。最近在大型语言模型方面的发展，例如ChatGPT，使得其他基金更容易加入竞争。虽然目前一个基金没有AI平台并不构成竞争劣势，但两年后将会是。

对于一个投资基金来说，交易来源和交易洞察是首先开始自动化的功能。然而，投资周期中还有许多其他功能可以通过这样的专有平台实现自动化。例如，整个投资生态系统的图模型可以用来向基金推荐投资者，向投资组合推荐个人等。同样的行业360度视角可以用来验证和更新基金的投资论点，运行投资组合模拟，并推荐投资组合调整的路径。此外，预测引擎可以用来预测交易的价格。基金可以使用他们的AI平台来推荐其投资组合公司的转型，支持它们使用自动化用例，并监控它们的进展。作为另一个例子，基金可以使用自动化机器来预测资产的正确退出时间，推荐潜在买家名单，并自动化供应商尽职调查过程。这些只是未来一些选定的用例。那些在内部为这样一个AI平台奠定正确基础的基金将在短期和长期内获得巨大优势。

The AI platform disruption has already started in the investment sector, now at the end of the fermentation phase of the S-curve, and beginning of the takeoff phase. Most major funds and many smaller funds have already started developing such in-house proprietary AI platforms. The recent developments in the large language models, e.g. ChatGPT, have made it easier for other funds to enter the game. While it is not a competitive disadvantage for a fund to not have an AI platform now, it will be in two years.

For an investment fund, deal sourcing and deal insights are the first functions to start automating. However, there are many other functions across the investment cycle that can be automated with such a proprietary platform. For example, the graph model of the entire investment ecosystem can be used to recommend investors to the funds, individuals to the portcos, etc. The same 360-degree view of the industry can be used for validating and updating the fund investment thesis, running a portfolio simulation, and recommending paths for portfolio correction. Moreover, the prediction engine can be used to predict the price of a deal. The funds can use their AI platform to recommend transformation for their portfolio companies, support them with automation use cases, and monitor their progress. For another example, the fund can use the automation machine to predict the correct exit time of an asset, recommend a list of potential buyers, and automate the vendor due diligence process. These are just some selected use cases for the future. Those funds that put the correct foundation for such an AI platform internally will garner large advantages in the short and long run.

Authors’ biography:

作者简介：

Mohammad Rasouli博士是投资领域人工智能和技术的思想领袖。他曾是湾区和纽约办公室的麦肯锡咨询公司顾问，负责管理私募股权客户的AI激活项目。他与前20大私募股权公司以及一些中小型PE公司合作，使用AI技术自动化他们的基金流程，并为他们的投资组合公司使用AI。在加入麦肯锡之前，Mohammad是斯坦福大学的AI研究员，在顶级AI会议上发表论文，获得450多次引用。他特别研究了推荐系统，并与公司紧密合作。他还在斯坦福大学商学院共同教授了“在线市场实证”课程。他在密歇根大学获得了电气和计算机工程博士学位，以及经济学硕士学位，论文委员会包括来自斯坦福、哈佛和麻省理工学院的教授，其中包括诺贝尔奖获得者。

Mohammad还曾是微软的工程师，并在初创企业中拥有经验，包括担任CTO和CEO职位。

Dr. Mohammad Rasouli is a thought leader in AI and technology for investment. He is an exMcKinsey consultant from the Bay Area and the New York offices managing AI activation projects for private equity clients. He has worked with top-20 private equities as well as a number of middle and small-size PEs to use AI technology for automating their fund processes as well as to use AI for their portfolio companies. Before joining McKinsey, Mohammad was an AI researcher at Stanford University publishing papers in top AI conferences with 450+ citations. He has particularly studied recommendation systems, working closely with companies. He has also co-taught “Empirics of Online Marketplaces” at Stanford Graduate School of Business. He finished his Ph.D. in Electrical and Computer Engineering, and his Masters in Economics from the University of Michigan working with a thesis committee of Stanford, Harvard, and MIT professors, including Nobel prize winners. Mohammad has also been a Microsoft engineer and has experience in startups including CTO and CEO roles.

Ravi Chiruvolu是一位科技愿景家，他拥有独特的视角，既是受过麻省理工学院培训的AI工程师，也是拥有25年以上顶级私营公司投资经验的投资者。在麻省理工学院和美国国家航空航天局（NASA）期间，他帮助创建了专家系统和复杂任务的视觉显示系统（在输入有限的情况下远程控制机器人），因此他对将AI平台应用于投资以创造可持续竞争优势的机会有着直观的理解。Ravi拥有麻省理工学院的学士和硕士学位，在多个NASA设施工作过，曾在麦肯锡担任商业分析师，在Ameritech担任企业战略师，以及两家投资基金（Alta Partners和Charter Ventures）的普通合伙人和管理合伙人。Ravi拥有哈佛商学院的MBA学位，并且是美国陆军的退役中尉。

Ravi Chiruvolu is a technology visionary and provides a unique perspective of being a MIT trained AI engineer and a top quartile private company investor for 25+ years. While at MIT and NASA, he helped create expert systems and a visual display systems for complex tasks (remotely controlling robots despite limited inputs), so maximizing the opportunity to apply AI platforms to investing to create sustainable competitive advantage is something he intuitively understands. Ravi has a BS and MS from MIT, worked at multiple NASA facilities, was a Business Analyst at McKinsey, a corporate strategist at Ameritech, and a General Partner and Managing Partner of two Investment Funds (Alta Partners and Charter Ventures). Ravi has an MBA from Harvard Business School and is a Retired First Lieutenant in the US Army.

Ali Risheh是一位AI研究员和工程师。他的研究领域涉及大型语言模型和神经网络，特别是图卷积神经网络。他曾与来自斯坦福大学、弗吉尼亚理工学院和多伦多大学的研究人员合作。他在加利福尼亚州立大学洛杉矶分校攻读了计算机科学的硕士学位。Ali曾是一位敏捷工程师，也是机器学习初创公司的首席技术官（CTO）和联合创始人。

Ali Risheh is an AI researcher and engineer. His research has been in the field of large language models and neural networks, particularly graph convolutional neural networks. He has collaborated with researchers from Stanford University, Virginia Tech, and the University of Toronto. He has studied MSc in Computer Science at California State Los Angeles University. Ali has been a prompt engineer, also CTO and co-founder of machine learning startups.

References

[1] McKinsey Global Private Markets Review: Private markets turn down the volume. McKinsey and Company, 2023.

[2] Tech providers 2025 special report. Gartner, 2021.

[3] 2023 global asset & wealth management survey. PwC, 2023.

[4] The Impact Of AI In Private Capital. Forbes, Aug 2023.

[5] LPs expect expanding role of AI in PE - Coller Capital Barometer. Private Equity International News and Analytics, June 2023.

[6] Alternative data: The new oil for the digital economy? Lowenstein Sandler, 2022.

[7] Finding the informational advantage in private equity. BlackRock, 2023.

[8] Andre Retterath. Data-driven VC Landscape. Early bird Venture Capital, 2023.

[9] EQT’s Alexandra Lutz: "Motherbrain will be the backbone of investment life cycle". PE HUB EUROPE, 2022. [10] How AI is transforming investing. BlackRock Systematic Investing, 2023.

[11] Are you ready for private equity GPT revolution? Private Equity International Editor’s View, May 2023. [12] Krafft P.; Moro E.; Pentland A. Adjodah D.; Leng Y.; Chong S. Accuracy-risk trade-off due to social learning in crowd-sourced financial predictions. Entropy, 2021.

[13] Google DeepMind’s Co-founder: My new Turing test would see if AI can make $1 million. MIT Technology Review, June 2023.

[14] Cao L.; von Ehrenheim V.; Krakowski S.; Li X.; & Lutz A. Using deep learning to find the next unicorn: A practical synthesis on optimization target, feature selection, data split and evaluation strategy. IJCAI Workshop of Multimodal AI For Financial Forecasting (MuFFin), ACL Anthology, 2023.

[15] Chen M.; Wang C.; Qin C.; Xu T.; Ma J.; Chen E.; Xiong H. A trend-aware investment target recommendation system with heterogeneous graph. International Joint Conference on Neural Networks (IJCNN), 2021.

[16] Catovic A.; Ferm A.; Ingelhag H. Cao L.; von Ehrenheim V.; Berghult A., Henje C.; Stahl R.; Wandborg J.; Stan S. A scalable and adaptive system to infer the industry sectors of companies: Prompt + model tuning of generative language models. EQT Motherbrain, 2023.

[17] Dellermann D.; Lipusch N.; Ebel P.; Popp K.; Leimeister J. Finding the unicorn: Predicting early stage startup success through a hybrid intelligence method. arxiv, 2021.

[18] Mike A. Transformers, llms, and the (r)evolution of the moonfire tech stack. Moonfire Pulse, 2023.

[19] Krakovna V.; Doshi-Velez F. Increasing the interpretability of recurrent neural networks using hidden markov models. arxiv, 2016.