Multifundamentals: Unified Multi-Anchor Models

• Multifundamentals are a family of models that replace a single foundational object with multiple interacting components, as seen in algebraic combinatorics and financial risk modeling.
• They employ techniques ranging from K-theoretic lifts and operator constructions to variance-preserving decompositions, ensuring structural versatility and robust representations.
• Applications span from enhanced portfolio theory and deep factor investing to integrating heterogeneous financial data in multimodal models for actionable insights.

In the cited literature, multifundamentals does not designate a single standardized object. It appears instead as a family of domain-specific constructions built from multiple simultaneous “fundamental” components: in algebraic combinatorics, the multifundamental quasisymmetric functions $\overline{F}_a$ are $K$-theoretic lifts of ordinary fundamental quasisymmetric functions; in quantitative finance, the term is used for models that combine multiple firm characteristics or reduce a large asset universe to a smaller set of tradable funds; and in financial AI it broadens further to unified models over fundamental, market, macroeconomic, and alternative data (Pierson, 15 Aug 2025, Kakushadze et al., 2016, Kardaras et al., 2022, Wei et al., 2022, Yanglet et al., 15 May 2025).

1. Domain scope and recurring idea

Across the sources, the common pattern is the replacement of a single foundational object by a structured collection. In one case the collection is indexed by a strong composition and encoded by set-valued combinatorics; in another it is a vector of factor loadings or tradable funds; in another it is an interleaving of heterogeneous financial modalities. This suggests a family resemblance rather than a unified theory: the same label marks settings where one studies simultaneous structure, coupled components, or multi-anchor representations instead of a single basis, factor, or modality (Pierson, 15 Aug 2025, Kakushadze et al., 2016, Yanglet et al., 15 May 2025).

Domain	Core referent	Representative source
Algebraic combinatorics	Multifundamental quasisymmetric functions $\overline{F}_a$	(Pierson, 15 Aug 2025)
Equity risk modeling	Custom multifactor models using arbitrary characteristics with variance-preserving construction	(Kakushadze et al., 2016)
Continuous-time portfolio theory	Fund models where the GOP lies in a low-dimensional fund span	(Kardaras et al., 2022)
Deep factor investing	Nonlinear deep factors learned from multiple stock characteristics	(Wei et al., 2022)
Financial foundation models	Interleaved multimodal financial data, including fundamental data	(Yanglet et al., 15 May 2025)

2. Multifundamentals in algebraic combinatorics

In algebraic combinatorics, the relevant object is the multifundamental quasisymmetric function $\overline{F}_a$, introduced by Lam–Pylyavskyy as $\tilde L_a$ and indexed by a strong composition $a=a_1a_2\cdots a_\ell$. The associated partial-sum set is

$$\operatorname{set}(a):=\{a_1,\ a_1+a_2,\ \dots,\ a_1+\cdots+a_{\ell-1}\},$$

and $|a|:=a_1+\cdots+a_\ell$. The ordinary fundamental quasisymmetric function $F_a(x_1,\dots,x_n)$ is the generating function for weakly increasing sequences $1\le i_1\le \cdots \le i_{|a|}\le n$ with strict increases at positions in $K$0. The multifundamental $K$1 is its $K$2-theoretic lift, obtained by replacing singleton entries by set-valued entries (Pierson, 15 Aug 2025).

The defining combinatorial model uses weakly increasing sequences of sets

$K$3

with $K$4, and with strict increases $K$5 required for $K$6. The order is not inclusion but

$K$7

The generating series is

$K$8

The parameter $K$9 records the set-valued “excess”

$\overline{F}_a$0

Setting $\overline{F}_a$1 forces each $\overline{F}_a$2 to have size $\overline{F}_a$3, so that

$\overline{F}_a$4

This is the paper’s explicit reason for viewing $\overline{F}_a$5 as a $\overline{F}_a$6-theoretic analogue: ordinary objects are replaced by set-valued ones, and $\overline{F}_a$7 records the excess beyond the classical model (Pierson, 15 Aug 2025).

3. Operator-theoretic characterization and hierarchy

The main novelty of the 2025 paper is an operator construction that places multifundamentals into the same structural hierarchy as Schur polynomials, Demazure characters, Grothendieck polynomials, fundamental slides, and kaons. Hivert’s modified swap operator $\overline{F}_a$8 yields “fundamental” divided differences $\overline{F}_a$9 and $\overline{F}_a$0. The paper then defines $\overline{F}_a$1-analogues of Hivert’s operators,

$\overline{F}_a$2

which specialize at $\overline{F}_a$3 to Hivert’s operators (Pierson, 15 Aug 2025).

These operators satisfy Hecke-type relations, including

$\overline{F}_a$4

together with braid and commutation relations, so reduced-word definitions are well posed. A key local property is that if a monomial involves both $\overline{F}_a$5 and $\overline{F}_a$6, then $\overline{F}_a$7 fixes that monomial and $\overline{F}_a$8 annihilates it. This mirrors Hivert’s setup and underlies the inductive combinatorial proof (Pierson, 15 Aug 2025).

The central theorem identifies multifundamentals as images of monomials under the longest-word operator: $\overline{F}_a$9 and, more precisely in finite variables,

$\tilde L_a$0

Within the same operator package, the paper also recovers the fundamental glides $\tilde L_a$1 and the kaons $\tilde L_a$2, so multifundamentals occupy the “global” $\tilde L_a$3-position in the $\tilde L_a$4-Hivert hierarchy. The resulting basis is positive over $\tilde L_a$5, specializes to $\tilde L_a$6 at $\tilde L_a$7, and admits a finite-variable stability statement because $\tilde L_a$8 may be chosen independently of $\tilde L_a$9 (Pierson, 15 Aug 2025).

A representative example is $a=a_1a_2\cdots a_\ell$0 with $a=a_1a_2\cdots a_\ell$1, for which $a=a_1a_2\cdots a_\ell$2. The ordinary fundamental is

$a=a_1a_2\cdots a_\ell$3

while the multifundamental adds the set-valued terms

$a=a_1a_2\cdots a_\ell$4

The extra $a=a_1a_2\cdots a_\ell$5-terms encode precisely the excess from replacing singleton letters by set-valued letters (Pierson, 15 Aug 2025).

4. Financial meanings: characteristic-based risk models and fund spans

In quantitative finance, one meaning of multifundamentals is methodological rather than terminological: a model that incorporates multiple firm characteristics into a coherent covariance or growth framework. In the custom multifactor risk-model literature, the basic object is the factor-model decomposition

$a=a_1a_2\cdots a_\ell$6

with diagonal specific risk $a=a_1a_2\cdots a_\ell$7, factor loadings $a=a_1a_2\cdots a_\ell$8, and factor covariance $a=a_1a_2\cdots a_\ell$9. The paper stresses that a valid risk model should reproduce sample variances, $$\operatorname{set}(a):=\{a_1,\ a_1+a_2,\ \dots,\ a_1+\cdots+a_{\ell-1}\},$$0, and proposes a rescaled correlation-space construction

$$\operatorname{set}(a):=\{a_1,\ a_1+a_2,\ \dots,\ a_1+\cdots+a_{\ell-1}\},$$1

with rescaled loadings $$\operatorname{set}(a):=\{a_1,\ a_1+a_2,\ \dots,\ a_1+\cdots+a_{\ell-1}\},$$2 and $$\operatorname{set}(a):=\{a_1,\ a_1+a_2,\ \dots,\ a_1+\cdots+a_{\ell-1}\},$$3. The paper explicitly states that arbitrary characteristic vectors can be used as factor loadings, but should be inserted into this variance-preserving normalized-correlation construction rather than a naive covariance regression. It also distinguishes principal components, industry factors, and non-industry factors, and proposes the heterotic and Russian-doll constructions to obtain nonsingular factor covariance matrices at short horizons (Kakushadze et al., 2016).

The same paper’s closest link to multifundamentals is heterotic CAPM, where a characteristic $$\operatorname{set}(a):=\{a_1,\ a_1+a_2,\ \dots,\ a_1+\cdots+a_{\ell-1}\},$$4 is used as a within-subindustry weighting: $$\operatorname{set}(a):=\{a_1,\ a_1+a_2,\ \dots,\ a_1+\cdots+a_{\ell-1}\},$$5 This permits accounting, valuation, profitability, quality, growth, leverage, or similar signals to enter either as direct factor-loading columns or as within-industry weights. The paper also warns that centering and normalization matter because factor models are invariant to nonsingular linear transformations of columns of $$\operatorname{set}(a):=\{a_1,\ a_1+a_2,\ \dots,\ a_1+\cdots+a_{\ell-1}\},$$6, but not to arbitrary shifts $$\operatorname{set}(a):=\{a_1,\ a_1+a_2,\ \dots,\ a_1+\cdots+a_{\ell-1}\},$$7. Its empirical conclusion is cautious: appending a few style factors or principal components to a granular heterotic industry model adds little, whereas using a style signal as the within-subindustry weighting can improve some metrics, especially cents-per-share (Kakushadze et al., 2016).

A second financial meaning appears in continuous-time fund models. There, a market is described by a lower-dimensional collection of tradable funds $$\operatorname{set}(a):=\{a_1,\ a_1+a_2,\ \dots,\ a_1+\cdots+a_{\ell-1}\},$$8 such that asset returns satisfy

$$\operatorname{set}(a):=\{a_1,\ a_1+a_2,\ \dots,\ a_1+\cdots+a_{\ell-1}\},$$9

with the residuals $|a|:=a_1+\cdots+a_\ell$0 local martingales. Proposition 1 in that paper states the equivalence between this decomposition and the statement that the global growth-optimal portfolio $|a|:=a_1+\cdots+a_\ell$1 lies in the fund span,

$|a|:=a_1+\cdots+a_\ell$2

This yields a precise sense in which a high-dimensional market may possess a low-dimensional multifund structure: all priced return opportunities relevant for growth-optimal investing lie in the span of $|a|:=a_1+\cdots+a_\ell$3 tradable funds (Kardaras et al., 2022).

The paper’s central estimation result is that, under local frequentist estimation in a correct fund model, the expected instantaneous growth loss depends only on the number of funds: $|a|:=a_1+\cdots+a_\ell$4 Under Bayesian filtering, the growth loss is

$|a|:=a_1+\cdots+a_\ell$5

and in the special Gaussian updating case with $|a|:=a_1+\cdots+a_\ell$6, $|a|:=a_1+\cdots+a_\ell$7, this becomes

$|a|:=a_1+\cdots+a_\ell$8

The practical implication stated in the paper is that a correctly specified low-dimensional fund span reduces the economic cost of estimation and learning. The same article then proposes shrinkage of the filtered GOP estimate toward zero exposure, with the shrunk portfolio

$|a|:=a_1+\cdots+a_\ell$9

and reports that shrinkage gives a more stable estimate that more closely tracks growth potential than an unrestricted Bayesian estimate (Kardaras et al., 2022).

5. Learned nonlinear and multimodal financial multifundamentals

A more recent usage shifts from hand-crafted factors to learned latent factors. In the deep multi-factor model for factor investing, the raw input panel is

$F_a(x_1,\dots,x_n)$0

with $F_a(x_1,\dots,x_n)$1 factors drawn from the groups reversal, value, size, momentum, and quality. The model applies

$F_a(x_1,\dots,x_n)$2

then constructs an industry graph and an industry-neutral context

$F_a(x_1,\dots,x_n)$3

followed by a universe graph and universe-neutral context

$F_a(x_1,\dots,x_n)$4

The deep factor is then learned from the concatenation of original, industry-neutralized, and universe-neutralized contexts, and separate heads are trained for horizons $F_a(x_1,\dots,x_n)$5. An attention module approximately composes each learned factor from the original factors, and the paper reports that quality-group factors receive the largest attention for the $F_a(x_1,\dots,x_n)$6-day factor across CSI1000, CSI500, and CSI300 (Wei et al., 2022).

This learned-factor usage differs from classical multifactor modeling in two ways stated explicitly in the paper. First, the model is intended to learn nonlinear combinations and interactions among observed factors rather than fixed linear blends. Second, neutralization is implemented as a learned graph operation rather than a standard regression step. The evaluation protocol uses monthly rebalancing, an equal-weight long-only top-decile portfolio, and a transaction cost of $F_a(x_1,\dots,x_n)$7, with metrics $F_a(x_1,\dots,x_n)$8, ICIR, IR, and SR. On the reported Chinese universes, the model generally improves portfolio outcomes over linear regression, equal-weight combinations, a plain MLP, and a universe-graph baseline (Wei et al., 2022).

At a broader scale, Multimodal Financial Foundation Models generalize multifundamentals from “many stock characteristics” to “many financial data systems.” The defining claim is that MFFMs ingest interleaved multimodal financial data, including fundamental data, market data, data analytics, macroeconomic data, and alternative data such as natural language, audio, images, and video. The paper maps this to concrete sources: financial reports such as 10-Q, 10-K, DEF 14A, 8-K, earnings releases, annual reports, Zacks reports, and sell-side broker reports; XBRL filings; earnings conference calls; monetary policy conferences; climate data; financial news; charts; tables; and time series. It also discusses FinAgents such as search agents, tutor agents, an XBRL agent, and FinRL trading agents, together with retrieval-augmented and tool-augmented workflows (Yanglet et al., 15 May 2025).

This literature does not define a formal theory of multifundamentals, but it does broaden the term’s operational scope. A plausible implication is that the “fundamental” units are no longer only accounting or valuation variables; they include corporate filings, market micro/macro signals, governance and ESG information, analytical artifacts, and alternative data streams. In that sense, the MFFM literature treats finance as inherently multi-fundamental because useful reasoning must align text, tables, numbers, time series, charts, and tool outputs within one model or agentic system (Yanglet et al., 15 May 2025).

6. Related integrative constructions and comparative perspective

Related papers reinforce the same methodological move: complex systems are modeled through multiple simultaneous structural layers rather than one canonical representation. In financial network analysis, for example, dependence is represented as a multiplex network with four layers corresponding to linear, non-linear, tail, and partial correlations. The four dependence measures have highly correlated average levels over time, but their network structures differ strongly, and edge overlap drops sharply around periods of financial stress. This does not define multifundamentals directly, but it provides an analogous picture in which multiple dependence “fundamentals” coexist and become especially distinct in crises (Musmeci et al., 2016).

A computational analogue appears in the FAMLIES framework for dense linear and multilinear infrastructure. That paper argues for a vertically integrated software stack spanning matrices and tensors, with six “girders”: Consistent Application Programming Interface, Comprehensive Control of Communication, Algorithm Control and Exploration of the Algorithmic Space, Support of Legacy and Future Applications, Representative Functionality, and Community. The operations explicitly include LU with pivoting, Cholesky, Householder QR, tensor contractions, tensor contraction networks, tensor transposition or permutation, CP-ALS, and HOOI/HOSVD. Here the relevant pattern is again multiplicity of foundational components—linear and multilinear kernels, layouts, communication patterns, and control trees—combined inside one framework rather than treated as separate silos (Matthews et al., 8 Apr 2026).

Taken together, these literatures suggest that multifundamentals is best understood as a cross-domain label for structured plurality. In algebraic combinatorics it names a specific $F_a(x_1,\dots,x_n)$9-theoretic basis $1\le i_1\le \cdots \le i_{|a|}\le n$0. In finance it can denote characteristic-rich factor and risk architectures, low-dimensional tradable fund spans for growth-optimal investing, or nonlinear and multimodal systems that integrate many classes of financial information. In related methodological work, the same organizing principle appears as multilayer dependence, multilinear infrastructure, or multi-anchor valuation dynamics. The unifying theme is not a shared formalism, but a repeated replacement of one fundamental representation by several interacting ones (Pierson, 15 Aug 2025, Kakushadze et al., 2016, Kardaras et al., 2022, Wei et al., 2022, Yanglet et al., 15 May 2025, Musmeci et al., 2016, Matthews et al., 8 Apr 2026).