Total Portfolio Approach: Holistic Optimization

• Total Portfolio Approach is a holistic optimization method that jointly manages interdependent components across wealth accounts, funds, or projects.
• It incorporates dynamic decision models, including mental accounting, tracking-error governance, and multi-stage IT project scheduling.
• The approach emphasizes simultaneous allocation, timing, and cross-linkages to enhance performance and risk management across systems.

Searching arXiv for the cited papers to ground the article in the current record. arxiv_search: (Bayraktar et al., 7 Jun 2025)

The Total Portfolio Approach denotes a family of portfolio-management formulations in which optimization is performed over the entire decision set rather than over isolated sleeves, accounts, or projects. In the recent arXiv literature, this label appears in at least three technically distinct settings: a continuous-time, goal-based investment problem with mental accounting and costly transfers across wealth accounts (Bayraktar et al., 7 Jun 2025); an institutional asset-allocation framework in which Total Portfolio Approach and Strategic Asset Allocation are separated by the width and cyclicality of a tracking-error constraint (Alankar et al., 3 Mar 2026); and a multi-period optimization model for interdependent IT projects under budgetary and sequencing constraints (Pushkar et al., 2010). Across these settings, the common object is joint optimization under explicit cross-component interactions, though the state variables, constraints, and governance mechanisms differ materially.

1. Conceptual scope

The term is used to describe holistic optimization over a portfolio as a system. In the investment-goals setting, the system consists of multiple wealth accounts connected by transfers and managed jointly (Bayraktar et al., 7 Jun 2025). In the institutional-governance setting, the system is the entire fund, with the chief investment officer acting under a single total-fund tracking-error budget rather than sleeve-specific budgets (Alankar et al., 3 Mar 2026). In the IT-project setting, the system is a multi-stage portfolio of interdependent projects whose value depends jointly on discounted cash-flow benefits, option-like dependencies, and implementation timing (Pushkar et al., 2010).

Domain	Portfolio objects	Coordinating mechanism
Goal-based investing	$K+1$ wealth accounts	Inter-account transfers with mental costs
Institutional asset allocation	Entire fund relative to benchmark	Dynamic tracking-error budget
IT project selection	Projects across $N$ periods	Joint optimization under budgets and sequencing

The significance of this usage is that “portfolio” is not restricted to a static collection of assets. It can refer to sub-portfolios tied to goals, active bets around a benchmark, or projects distributed across periods. A plausible implication is that the Total Portfolio Approach is best understood not as a single model class, but as a design principle: optimize the whole constrained system rather than optimize compartments independently.

2. Goal-based continuous-time formulation with mental accounting

In "Goal-based portfolio selection with mental accounting" (Bayraktar et al., 7 Jun 2025), an investor has $K+1$ goals with targets $G_k$ at deadlines $T_k$, where

$0=T_0< T_1<\dots<T_{K+1}=T.$

The investor runs $K+1$ wealth accounts $X_k(t)$, each invested in a Black–Scholes market and interconnected by transfers with proportional mental costs $\lambda_k$ for transfers into account $k$ and $N$0 for transfers out of account $N$1. The market is driven by a $N$2-dimensional Brownian motion and a state factor $N$3, with factor and risky-asset dynamics

$N$4

$N$5

If $N$6 denotes the fraction invested in stocks for account $N$7, the wealth processes satisfy coupled SDEs. For the last, “fundamental,” account $N$8,

$N$9

while for $K+1$0,

$K+1$1

Here $K+1$2 are nondecreasing RCLL processes representing cumulative transfers in and out of account $K+1$3, and a no-bankruptcy constraint $K+1$4 holds.

The investor minimizes expected discounted shortfall penalties at the deadlines together with mental costs: $K+1$5 The formulation is goal-based because each target has its own account and deadline; it exhibits mental accounting because transfers across accounts incur explicit penalties. The model therefore departs from a single-omnibus-portfolio representation even though all accounts share the same market opportunity set.

3. HJB system, constrained viscosity solutions, and free boundaries

On each subinterval $K+1$6 with active accounts $K+1$7, the state is $K+1$8. The Hamiltonian is written as

$K+1$9

where the omitted term collects the usual drift and covariance contributions. Mental-cost gradient constraints induce a quasi-variational operator

$G_k$0

Right before a deadline $G_k$1, the value $G_k$2 must match either funding goal $G_k$3 or immediate transfer into the next regime $G_k$4, again under the same gradient constraints. On the last interval $G_k$5, only account $G_k$6 remains, so

$G_k$7

Because the state is constrained by $G_k$8 and the transfer conditions are two-sided gradient inequalities, the appropriate solution concept is a constrained viscosity subsolution/supersolution. The paper shows uniqueness through a tailored comparison proof in the spirit of Crandall–Ishii–Lions, while using the stochastic Perron method to construct upper and lower envelopes of stochastic supersolutions and subsolutions (Bayraktar et al., 7 Jun 2025). A strict classical subsolution $G_k$9 together with the envelopes $T_k$0 and $T_k$1 yields continuity and the comparison principle.

The state space on each time slice splits into a continuation region $T_k$2, a transfer-into-$T_k$3 region, and a transfer-out-of-$T_k$4 region. The associated free boundaries are nontrivial, time-dependent, and generally non-monotone: they can bulge or notch outwards. In the continuation region, the minimal-Hamiltonian argument gives the feedback rule

$T_k$5

Because $T_k$6 depends on all active accounts, the optimal allocation for one goal depends on the wealth levels of the others. When transfer constraints bind, the state jumps along a $T_k$7 or $T_k$8 direction in the $T_k$9-plane until the boundary is reached.

Numerically, the model is implemented by a finite-difference discretization of $0=T_0< T_1<\dots<T_{K+1}=T.$0, $0=T_0< T_1<\dots<T_{K+1}=T.$1, and $0=T_0< T_1<\dots<T_{K+1}=T.$2 on a rectangular grid, with a penalty method for the gradient constraints and policy iteration for $0=T_0< T_1<\dots<T_{K+1}=T.$3. The reported numerical findings are fourfold: the transfer regions exhibit bulges and notches; with two active goals, the optimal $0=T_0< T_1<\dots<T_{K+1}=T.$4 for fund 1 depends on fund 2’s level and vice versa; surplus in one account may be held in cash or left invested in stocks to hedge the other account’s risk, depending on correlation and mental costs; and transfers from an overfunded important goal to a less important one may be deferred until the former’s deadline approaches (Bayraktar et al., 7 Jun 2025).

4. Institutional Total Portfolio Approach as dynamic tracking-error governance

"Dynamic Tracking Error and the Total Portfolio Approach" (Alankar et al., 3 Mar 2026) treats Total Portfolio Approach and Strategic Asset Allocation as variants of the same mean–variance problem. In this formulation, Strategic Asset Allocation is characterized by fixed asset-class weights, tight static tracking-error budgets for each sleeve, periodic and mechanical rebalancing, and governance via constraints. Total Portfolio Approach is characterized by CIO discretion over the entire fund with a single total-fund tracking-error budget, the ability to shift capital across all risk sources and vary active risk over time, and governance via trust.

The paper’s key insight is that both approaches solve

$0=T_0< T_1<\dots<T_{K+1}=T.$5

and differ only in the choice of the governance cap $0=T_0< T_1<\dots<T_{K+1}=T.$6 and its responsiveness to drawdowns. The ex ante tracking error is

$0=T_0< T_1<\dots<T_{K+1}=T.$7

where $0=T_0< T_1<\dots<T_{K+1}=T.$8 denotes portfolio weights, $0=T_0< T_1<\dots<T_{K+1}=T.$9 benchmark weights, and $K+1$0 the covariance matrix of asset returns.

For a single active dimension with active weight $K+1$1 and active-return volatility $K+1$2,

$K+1$3

and the expected compound active return in period $K+1$4 is

$K+1$5

where $K+1$6 is the expected active return per unit $K+1$7. The unconstrained optimum is

$K+1$8

so under regime-switching and cap $K+1$9, the CIO chooses

$X_k(t)$0

The framework fully characterizes realized tracking error by three moments: its level, its volatility $X_k(t)$1, and its cyclicality measured by $X_k(t)$2. This replaces a categorical distinction between TPA and SAA with a continuous governance object, namely the tracking-error policy function.

5. Empirical results and governance implications in institutional portfolios

The empirical design uses U.S. equity and bond data from January 2000 to February 2026, with simulations from 2004 to 2026. The benchmark is a monthly rebalanced 70/30 SPX/AGG portfolio; the active bet is the SPX–AGG spread scaled daily using a 63-day rolling $X_k(t)$3; and the regime proxy is the 21-day moving average VIX, with low regime when VIX $X_k(t)$4 and high regime when VIX $X_k(t)$5 (Alankar et al., 3 Mar 2026). A static-TE portfolio uses a constant $X_k(t)$6 annualized tracking error, while a dynamic-TE portfolio targets $X_k(t)$7 when VIX $X_k(t)$8, $X_k(t)$9 when $\lambda_k$0, and $\lambda_k$1 when VIX $\lambda_k$2. The full constraint spectrum spans 11 caps from $\lambda_k$3 to $\lambda_k$4.

Several quantitative findings are central. Across $\lambda_k$5 from $\lambda_k$6 to $\lambda_k$7, Sharpe ratios range from $\lambda_k$8 to $\lambda_k$9, statistically indistinguishable within the reported sample uncertainty, and the CAGR/Max DD ratio is likewise flat at approximately $k$0. By contrast, realized tracking-error volatility varies by about $k$1: $k$2 is approximately $k$3 at $k$4 and $k$5 at $k$6. The static-TE $k$7 portfolio has $k$8, whereas the dynamic-TE portfolio has $k$9 (Alankar et al., 3 Mar 2026).

The dynamic specification also outperforms the static $N$00 specification in CAGR, with $N$01 versus $N$02, while exhibiting annualized volatility of $N$03 versus $N$04, and essentially the same CAGR/Max DD ratio, $N$05 versus $N$06. The paper attributes part of this to a Jensen’s-inequality “volatility bonus,”

$N$07

The crisis-period opportunity set is summarized by the $N$08-premium. Forward SPX returns sorted by VIX quintiles from 1990 to 2026 imply a $N$09 percentage-point annualized spread at the one-month horizon between the highest- and lowest-VIX quintiles, with smaller but still positive spreads at 3-, 6-, and 12-month horizons. The “regret” of forced de-risking is also quantified: at GFC and COVID drawdown troughs, switching from 70/30 to 30/70 costs approximately $N$10–$N$11 percentage points over the next 12 months; in the 2022 rate shock, the same pattern appears with a smaller peak/trough and approximately $N$12 percentage points of 12-month regret (Alankar et al., 3 Mar 2026).

These results motivate a governance program centered on the tracking-error policy function. The paper recommends board reporting on the three moments of realized tracking error, evaluation horizons of at least a full market cycle, clawback provisions for tail-loss episodes, a cyclicality KPI based on positive $N$13 when opportunities peak, and documentation of triggers for risk addition or removal. It also states that private-asset allocations consume tracking-error budget without contributing to dynamic alpha, and that the net budget for liquid strategies should be large enough, specifically greater than $N$14, to permit crisis-period expansions. A further caution is explicit: a wide cap without a dynamic process yields static concentration, whereas the value attributed to TPA lies in time-varying active risk rather than merely large static deviations (Alankar et al., 3 Mar 2026).

6. Total Portfolio Approach in multi-period IT project optimization

In "A Metaheuristic Approach for IT Projects Portfolio Optimization" (Pushkar et al., 2010), the Total Portfolio Approach is applied to a multi-stage portfolio of IT projects. The model uses the project index set $N$15, implementation periods $N$16, period budgets $N$17, present-value project costs $N$18, discounted cash-flow benefits $N$19, dependency sets $N$20, dependency coefficients $N$21, cardinality bounds $N$22 and $N$23, and binary decision variables $N$24 indicating whether project $N$25 is implemented in period $N$26.

The objective maximizes portfolio value as the sum of discounted cash-flow contributions and real-option uplifts: $N$27 The constraints are: each project is implemented exactly once,

$N$28

period-wise budget feasibility,

$N$29

period-wise cardinality limits,

$N$30

and sequencing constraints for total dependencies,

$N$31

Partial dependencies with $N$32 do not introduce additional constraints; their effect enters through the objective.

The solution method is a genetic algorithm. A chromosome is a binary string of length $N$33, partitioned into $N$34 blocks of length $N$35, with block $N$36 encoding the projects implemented in period $N$37. Feasible chromosomes must satisfy the assignment and cardinality constraints. Initialization is random followed by repair; fitness is $N$38, with infeasible chromosomes penalized in proportion to budget shortfall or period overflow; selection is tournament selection of size 2 or roulette-wheel; crossover is one-point or two-point with typical probability $N$39; mutation flips each bit with typical probability $N$40, followed by repair; and termination occurs after a maximum generation count or when fitness improvement becomes negligible (Pushkar et al., 2010).

The illustrative computational experiment uses $N$41 projects over $N$42 periods, with budgets $N$43, $N$44, $N$45, and cardinalities $N$46, $N$47 for all $N$48. The best-found objective value is $N$49, mean population fitness at termination is approximately $N$50, convergence is reached near generation $N$51, and computation time is approximately $N$52–$N$53 seconds per run in MATLAB GA on a standard desktop. The reported solution structure places projects with highest out-degree in the dependency graph early, mixed dependency-sensitive projects in the middle, and primarily independent or low-option-value projects late (Pushkar et al., 2010).

In this setting, the Total Portfolio Approach refers to holistic optimization over projects and periods jointly. The model realizes trade-offs between current payoffs and future flexibility, embeds both strict and soft dependencies, and treats stage-wise funding as part of the optimization problem rather than as an external scheduling afterthought.

7. Unifying themes, limits, and common misconceptions

Several recurrent themes emerge across these formulations. First, the approach is holistic rather than sleeve-by-sleeve. In the mental-accounting model, multiple accounts share the same market but are managed jointly through transfers and coupled controls (Bayraktar et al., 7 Jun 2025). In the institutional model, the entire fund is governed through a single tracking-error budget and its dynamics rather than through static sleeve limits (Alankar et al., 3 Mar 2026). In the IT-project model, budgets, sequencing, and dependencies are optimized jointly across all projects and periods (Pushkar et al., 2010).

Second, cross-component interactions are central. In the goal-based model, the optimal risky allocation for one account depends on the wealth of other accounts, and transfer boundaries create coupled free-boundary geometry. In the institutional model, realized tracking error is evaluated by its level, volatility, and cyclicality, so governance concerns focus on the time path of active risk rather than merely its average level. In the IT-project model, project value includes dependency-driven option uplift, so one project’s scheduling changes another project’s effective contribution.

Third, the literature rejects several simplified interpretations. One simplification is that Total Portfolio Approach must mean a single undifferentiated portfolio; the goal-based mental-accounting paper states the opposite, namely that mental accounting leads to a family of sub-portfolios that are managed jointly rather than collapsed into an omnibus account (Bayraktar et al., 7 Jun 2025). A second simplification is that TPA and SAA are categorically distinct investment doctrines; the tracking-error paper argues that the only true governance difference is the width and cyclicality of the permitted tracking error (Alankar et al., 3 Mar 2026). A third simplification is that total-portfolio thinking is equivalent to ranking projects by stand-alone value; the IT-project formulation instead optimizes a combinatorial object with budgets, sequence dependence, and real-option effects (Pushkar et al., 2010).

A plausible synthesis is that the Total Portfolio Approach is best viewed as a system-level optimization doctrine whose concrete mathematical form depends on the domain. In continuous-time finance it appears as a constrained HJB problem with free boundaries; in institutional governance it appears as a dynamic tracking-error policy function; in project selection it appears as a constrained integer program solved metaheuristically. What is invariant is the insistence that allocation, timing, and cross-linkages be solved jointly rather than partitioned into locally optimized components.