No returns: Mega-IPOs and the promise of AI as economic measurement

The growing significance of Mega-IPOs

The emergence of exceptionally large initial public offerings is drawing renewed attention to the relationship between private capital, public markets and the capacity of institutional investors to absorb new financial exposures. Much of the commentary has focused on the valuations commanded by newly listed companies, the liquidity available to support their issuance and the investment opportunities they present. Yet beneath these familiar concerns lies a more fundamental development in financial markets. Companies are increasingly able to accumulate substantial economic value, operational complexity and financing requirements before establishing a meaningful public trading history, creating a growing population of financially significant assets whose underlying risks are difficult to assess using conventional market-based information.

The scale of this development is increasingly difficult to ignore. According to EY's Global IPO analysis, 1,293 companies raised approximately $171.8 billion through public offerings in 2025, representing a 39% increase in proceeds despite relatively little change in the number of listings. The concentration of capital within larger transactions is an important feature of this recovery, illustrating how a relatively small number of companies can account for a disproportionate share of public-market issuance. These offerings also arrive against the backdrop of a much larger accumulation of capital within private markets, where companies and investment vehicles have continued to expand outside the traditional structures of publicly traded securities.

McKinsey's 2026 Global Private Markets Report estimates private-capital assets under management at approximately $24 trillion in 2025, compared with around $14 trillion in 2020, using a definition that encompasses traditional funds alongside separately managed accounts, co-investments and other investment structures. Meanwhile, Bain's 2026 Global Private Equity Report identifies approximately 32,000 unsold buyout-backed companies representing $3.8 trillion in unrealised value. Although these measures describe different dimensions of private-market activity, together they illustrate the considerable stock of investments whose valuation, financing and eventual liquidity depend on processes that differ substantially from those governing continuously traded public securities.

For portfolio managers, the consequences extend well beyond the pricing of individual IPOs. A newly listed company may enter the investment universe with a multibillion-dollar market capitalisation but limited information about its behaviour under different financial conditions, while institutional investors may already hold related economic exposures through venture funds, private equity, credit arrangements or other companies within the same commercial ecosystem. The transition from private to public ownership therefore reveals a wider challenge in portfolio management: financial assets increasingly cross the boundaries between investment structures faster than their underlying risks can be measured and reconciled within established analytical frameworks.

From market transactions to portfolio risk

The immediate risks associated with Mega-IPOs are familiar to practitioners. Large offerings introduce demands on investor liquidity, create uncertainty around pricing and allocation, and may produce substantial changes in benchmark composition and portfolio concentration. Their market impact also depends on financing conditions, institutional positioning and the capacity of intermediaries to support trading around issuance. A company's headline valuation provides only a partial indication of these effects, since the actual capital absorbed depends on the shares offered, the investable float, the participation of existing shareholders and the financing arrangements employed by investors.

These considerations become more consequential when large offerings occur alongside broader changes in interest rates, investor risk appetite and access to credit. Growth-oriented companies with substantial future financing needs are exposed to changes in discount rates and refinancing conditions, while institutions acquiring their shares may need to rebalance existing portfolios or liquidate other investments. The resulting risks are therefore connected through both asset prices and funding arrangements, linking conventional systematic exposures with the possibility of financial contagion when leverage, collateral requirements and constrained liquidity amplify market movements.

At the institutional portfolio level, these events expose a difficult interaction between investment opportunity and the information required to evaluate it. A portfolio manager may need to establish a position before reliable estimates of volatility, correlation, liquidity and factor exposure are available, while the investment's economic characteristics may already overlap considerably with existing holdings. Index inclusion, changing free float and subsequent share sales create further adjustments in portfolio weights and risk contributions, even where the underlying company remains fundamentally unchanged.

The difficulty is particularly apparent in performance evaluation. Returns attributable to obtaining a primary-market allocation, purchasing shares after listing, timing an index rebalance or exercising investment judgment are economically distinct, although conventional attribution methods may not fully distinguish them. The same investment can generate markedly different results depending on allocation access, execution prices and the timing of benchmark recognition. Mega-IPOs therefore expose a gap between the statistical representation of an investment within a portfolio and the economic process through which that investment becomes available, financed, held and eventually realised.

The convergence of public and private markets

The significance of these challenges becomes clearer when placed within the broader transformation of institutional portfolios. Private equity, venture capital, private credit, infrastructure and other alternative investments have become established components of asset allocation, while institutional investors increasingly participate through combinations of direct ownership, pooled funds, co-investments and publicly traded instruments. These arrangements provide access to a wider range of economic opportunities, but they also introduce substantial differences in valuation, contractual rights, information availability and liquidity.

The distinction between public and private ownership is particularly important because an asset's economic exposure does not depend exclusively on its listing status. A private infrastructure company and a publicly traded infrastructure operator may respond to similar changes in financing costs, energy prices and customer demand, while exhibiting very different reported return patterns because their valuation mechanisms differ. Similarly, investments classified across venture capital, public technology equity and private credit may share concentrated exposure to the same underlying growth assumptions or commercial relationships. Asset-class diversification can consequently conceal economic concentration when portfolios are evaluated primarily through reported returns and institutional classifications.

This creates an increasingly important role for the Total Portfolio Approach, in which investment decisions are evaluated in relation to the economic exposures, liquidity requirements and objectives of the entire institution rather than through isolated asset-class allocations. The development reflects a broader recognition that portfolios are not simply collections of securities with observable returns, but combinations of ownership rights, contractual cash flows, financing obligations and contingent claims whose economic characteristics evolve through time.

The convergence of public and private markets consequently changes the informational foundation of portfolio management. Institutions must compare assets whose values are established through continuous market trading with others measured through appraisals, negotiated transactions, financing rounds or discounted cash-flow models. They must also incorporate investments whose underlying risks are visible through operating information, contractual commitments or economic relationships rather than through an established history of market prices. This increasingly heterogeneous universe challenges the assumption that the information required for portfolio construction can be represented consistently through conventional return and covariance estimates.

The measurement problem within portfolio theory

Modern Portfolio Theory provides a natural starting point for understanding these difficulties. In its familiar mean–variance formulation, portfolio construction involves selecting weights according to expected returns, covariances and investor preferences for risk:

$$\max_w \left(w^\top\mu-\frac{\lambda}{2}w^\top\Sigma w\right)$$

The framework remains a powerful way of expressing the trade-off between expected return and portfolio risk. Its practical application, however, depends on obtaining credible estimates of the quantities entering the optimisation and on translating the resulting weights into an implementable portfolio. As the investable universe becomes more heterogeneous, the production of these estimates becomes increasingly difficult, particularly when assets have sparse histories, inconsistent valuation conventions or significant restrictions on trading.

The distinction between reported returns and underlying economic risk is central. Private-market valuations frequently incorporate appraisal judgments and information that arrives less continuously than public-market prices, creating differences in the timing and responsiveness of reported returns. Getmansky, Lo and Makarov (2004) developed an influential analysis of return smoothing and serial correlation in hedge funds, establishing a framework for understanding how reported performance can understate economically relevant risk. For portfolios combining liquid and illiquid investments, similar measurement issues complicate estimates of volatility, correlation and diversification, particularly when different holdings reflect economic developments at different speeds.

Valuation uncertainty also arises from the financial structure of the investments themselves. Gornall and Strebulaev (2020) demonstrate how differences in preferred-share rights can materially alter the economic interpretation of venture-capital valuations. Their analysis illustrates a broader point: assets that appear comparable through headline valuations or accounting classifications may represent substantially different claims on future cash flows. Accurate portfolio measurement consequently requires an understanding of the mechanisms determining those cash flows, including leverage, seniority, contractual protections and embedded optionality.

The quantitative finance literature has developed important responses to uncertainty in portfolio inputs. Covariance shrinkage, Bayesian estimation, factor models and robust optimisation provide methods for stabilising estimates and recognising uncertainty in expected returns and risk parameters. Goldfarb and Iyengar (2003), for example, formalised portfolio optimisation under uncertain model parameters, demonstrating how investment decisions can be made across sets of plausible estimates rather than a single assumed distribution. These methods improve the treatment of uncertainty within portfolio construction, but their effectiveness ultimately depends on the economic content of the information being modelled.

This establishes a deeper distinction between improving portfolio optimisation and improving portfolio measurement. An increasingly sophisticated optimiser can make better use of estimated returns, correlations and constraints, while remaining dependent on the reliability of those underlying quantities. When those quantities are unavailable or poorly represented by historical prices, the central challenge moves upstream, from determining the optimal allocation to establishing what can be known about the economic characteristics of the assets themselves.

The economics of investment complexity

The convergence of public and private markets creates an increasingly demanding information problem for institutional investors. Evaluating heterogeneous assets requires the interpretation of financial statements, contractual rights, operating conditions and financing arrangements, often through specialised methods that are difficult to standardise across large investment universes. The resulting dependence on scarce professional expertise places the production of reliable financial measurements under growing economic pressure, particularly in an asset-management industry facing persistent fee compression and rising technology costs. Van Nieuwerburgh and Veldkamp (2010) establish a theoretical connection between costly information acquisition and portfolio choice, demonstrating how the allocation of attention and information influences diversification. As financial complexity increases, the effective investment opportunity set becomes constrained not only by capital and liquidity, but also by an institution's ability to measure and compare economic exposures consistently. Scalable financial measurement consequently becomes an important determinant of both portfolio quality and the economics of investment management.

The statistical and economic limits of financial machine learning

The difficulty facing financial machine learning is rooted in the relationship between the size of the available sample and the space of plausible hypotheses. Even several decades of daily asset returns provide only a few thousand observations, with substantially fewer independent observations of business cycles, financial crises, technological transitions and monetary regimes. Against this comparatively limited statistical evidence, modern learning architectures can represent an enormous range of possible relationships. Statistical learning theory makes clear that generalisation depends on the relationship between hypothesis complexity, sample information and the assumptions governing the underlying distribution, rather than predictive fit alone. Finance compounds this difficulty because market participants learn and adapt, while innovation, regulation and changing economic incentives continually reshape the data-generating process. The adaptive-market perspective of Lo (2004) and the broader traditions of behavioural finance, evolutionary game theory and agent-based modelling provide economic foundations for understanding this endogenous nonstationarity.

These conditions motivate a more substantive role for economic theory within statistical learning. In both the physical and social sciences, mechanistic interpretation provides a means of distinguishing plausible explanations from statistical relationships that happen to fit finite observations. Economic mechanisms constrain the hypothesis space through accounting identities, contractual payoffs, financing relationships, behavioural incentives and equilibrium conditions, while empirical evidence determines which mechanisms remain consistent with observed outcomes. In Bayesian terms, learning becomes the interaction between the likelihood assigned by the data and the prior plausibility of competing economic mechanisms. The purpose of mechanistic interpretation is therefore not merely to explain a fitted model after estimation, but to discipline what can be learned from limited observations, while establishing the economic conditions under which those relationships may persist.

This perspective also changes the meaning of interpretability. Conventional distinctions between local feature attribution, such as SHAP, and global model explanation concern how a fitted function can be inspected. For economically motivated representation learning, interpretation should instead begin with the training objective, the properties of the representations it induces and their relationship to identifiable economic states. Self-supervised methods such as DINO (Caron et al., 2021) illustrate how meaningful representations can emerge without explicit target labels, while DINOv2 (Oquab et al., 2023) extends this approach to general-purpose visual features. Their relevance to financial measurement lies less in explaining individual pixels or attention weights than in understanding the invariances induced by self-supervised learning, the economic phenomena encoded by the resulting representations and the market environments across which those measurements remain stable. The interpretability requirement is thus relocated from the internal mechanics of every model component to the economic meaning, validity and behaviour of the quantities being measured.

The research frontier: Economically structured representation learning

A promising research direction lies in combining self-supervised representation learning, alternative data and structural economic modelling to estimate otherwise poorly observed financial states. Rather than learning arbitrary predictors of asset returns, such systems would identify representations of operating capacity, demand, financing conditions, contractual exposures and other economic quantities from heterogeneous observations, with economic theory informing their interpretation and relationships. Probabilistic inference could then distinguish uncertainty in the underlying states from uncertainty about the measurements themselves, while mechanism-based constraints provide a basis for evaluating whether learned relationships remain economically coherent across changing market environments.

The scientific challenge is to establish which representations emerge from different learning objectives, how they correspond to economically meaningful quantities and which structural or behavioural conditions govern their persistence. Such research connects advances in representation learning with asset pricing, econometrics and economic measurement, providing a foundation for portfolio construction that is less dependent on historical return observations and more directly informed by the economic processes generating financial risk.

AI as economic measurement

An economically motivated measurement framework begins with the quantities institutional investors need to understand rather than the historical variables most readily available for prediction. Across public equity, private credit, venture capital and infrastructure, these quantities may include operating capacity, customer demand, profitability, financing requirements, contractual obligations, asset coverage, liquidity and sensitivity to common economic shocks. Their values are often only partially observed, although evidence about them is distributed across financial statements, transaction records, legal agreements, operating data and alternative information sources.

AI provides a means of combining these observations within models of underlying economic states. Rather than interpreting each data source as an independent predictor of asset prices, a measurement framework treats it as evidence about a defined financial or economic quantity. Information on electricity consumption may inform estimates of infrastructure utilisation, for example, while customer contracts provide evidence about revenue commitments and financing documents establish the obligations governing future cash flows. The economic interpretation of each observation is established through its relationship to the underlying activity or contractual claim being measured.

A general representation can be developed through a probabilistic measurement model:

$$x_{i,t}^{(k)}=h_k(z_{i,t},\theta_k)+\epsilon_{i,t}^{(k)}$$

where $x_{i,t}^{(k)}$ represents evidence from source $k$, $z_{i,t}$ is an underlying economic state, $h_k$ describes the relationship between that state and the observed evidence, and $\epsilon_{i,t}^{(k)}$ captures measurement error. The latent state may describe quantities such as demand, productive capacity, leverage or operating profitability, while additional economic relationships connect these states to future cash flows and the value of different financial claims.

This formulation creates an important distinction between a statistical representation and an economically interpretable one. Latent variables are useful for financial analysis when their meaning is established through accounting relationships, contractual structures or identifiable economic mechanisms. A model can consequently represent the relationship between operating conditions and investment outcomes in terms familiar to practitioners, preserving the distinction between an underlying business and the financial claims issued against it.

The approach also creates a natural role for alternative data. Hassan and co-authors (2019) demonstrate how earnings-call text can be used to construct firm-level measures of political risk, establishing a method for extracting economically interpretable information from corporate language. In a different setting, Rolf and co-authors (2021) show how satellite imagery can support the estimation of socioeconomic characteristics across different measurement tasks. Such work illustrates the potential to extend economic observation beyond conventional financial reports and historical market prices.

The next step is to integrate these measurements with financial mechanisms. For an infrastructure investment, observations about construction, operating capacity, utilisation and contractual revenue can inform estimates of cash generation, while financing agreements determine how that cash is distributed among creditors and equity holders. The resulting model can connect changes in economic conditions to asset-specific outcomes, allowing exposures across different financial structures to be compared through common underlying drivers.

This has particular value for portfolios containing both public and private claims on related economic activities. Rather than treating their reported returns as the principal measure of similarity, an institution could estimate their exposure to shared economic states and examine how those exposures interact under different scenarios. Portfolio diversification would then be evaluated through the economic relationships generating cash flows and losses, supported by market information where available.

Uncertainty becomes an integral part of this measurement process. A probabilistic framework can represent uncertainty associated with incomplete observations, model parameters, underlying economic states and future developments, carrying those uncertainties through to asset-level and portfolio-level outcomes. Financial measurements could consequently be expressed as distributions or ranges linked to identifiable economic assumptions, enabling risk managers to distinguish exposures supported by extensive evidence from those dependent on limited or conflicting information.

This is a significant conceptual departure from treating uncertainty primarily as the historical volatility of observed returns. It recognises uncertainty as a characteristic of the institution's knowledge about an investment, as well as a characteristic of the investment's economic outcomes. For assets with limited public histories, this distinction offers a means of constructing risk estimates from economically relevant evidence while preserving the limitations of what can presently be inferred.

The research frontier and the foundations of validation

Several developments in quantitative finance provide foundations for this approach. Chen, Pelger and Zhu (2024) integrate no-arbitrage restrictions into deep-learning asset pricing, demonstrating how economic structure can inform flexible statistical estimation. Related research on machine-learning stochastic discount factors seeks to combine nonlinear representations with financial pricing relationships, while work in probabilistic inference and state-space modelling offers methods for estimating unobserved quantities from incomplete evidence. Together, these developments point toward an increasingly productive intersection of financial economics, econometrics and machine learning, in which economic theory shapes both what is measured and how the resulting estimates are interpreted.

The principal research opportunity lies in integrating these approaches into a coherent framework for heterogeneous financial assets. Such a framework would require economically defined latent states, measurement models capable of incorporating structured and unstructured evidence, explicit treatment of information quality, and financial mechanisms that translate the estimated states into cash-flow and risk distributions. The resulting measurements could support portfolio construction across instruments with different ownership rights, liquidity arrangements and observation histories, while maintaining a common representation of the economic factors to which they are exposed.

Its value can be evaluated empirically. Conventional return-based models, fundamental factor models and economically structured AI measurement systems could be compared using identical investment universes and historically available information. Measurement quality would be assessed through subsequent disclosures, cash-flow realisations, defaults and transaction outcomes, while portfolio applications would be evaluated through changes in estimated concentration, risk calibration, liquidity management, implementation costs and risk-adjusted performance. The decisive measure of progress would be the improvement in economically achievable investment decisions relative to the costs of obtaining and maintaining the information.

The distinction between statistical validity and economic usefulness is particularly relevant to institutional governance. Financial professionals require models whose underlying measurements can be understood, independently challenged and evaluated against their intended applications. A framework built around explicitly defined economic quantities creates a basis for tracing estimates to their supporting evidence, identifying the assumptions governing their interpretation and understanding how uncertainty propagates into portfolio decisions. These properties connect naturally with the principles of diligence, reasonable basis and model validation embedded in professional investment practice and financial supervision.

The longer-term contribution could extend beyond improving individual valuation or risk models. A common economic measurement framework would allow institutions to organise information across asset classes according to the underlying drivers of financial outcomes, rather than relying exclusively on reporting conventions, historical prices or institutional product classifications. It would also establish a more systematic basis for determining the value of additional information, directing analytical resources toward investments and exposures where better measurement can materially improve portfolio decisions. In an industry facing growing investment complexity and persistent cost pressure, that combination of analytical consistency and economic scalability is especially consequential.

Conclusion

Mega-IPOs provide a visible expression of a broader transformation in financial markets, as substantial economic value accumulates within increasingly diverse forms of ownership, financing and investment. The convergence of public and private markets is expanding the range of opportunities available to institutional investors while exposing the limitations of financial measurements developed primarily around observable market returns. These limitations affect not only portfolio optimisation and risk management, but also the cost, capacity and consistency of the analytical processes through which investment decisions are made.

The resulting opportunity for artificial intelligence is more fundamental than improvements in financial prediction. Economically structured AI, supported by alternative data and probabilistic inference, offers a means of extending financial measurement to underlying economic states that are imperfectly captured by traditional market information. By connecting heterogeneous evidence to identifiable economic mechanisms and carrying uncertainty through to portfolio outcomes, this approach could strengthen the informational foundations of investment management.

The important frontier is therefore the development of a more comprehensive science of financial measurement, capable of representing economic exposure across public and private assets with greater consistency, interpretability and quantitative discipline. Its significance lies in shifting the foundations of portfolio analysis from the properties of observed financial returns toward the economic processes that generate them, and in making those processes measurable at a scale compatible with the changing structure of institutional finance.