Module 73 — Portfolio Construction & Optimization

How to build portfolios that are genuinely optimal — not just optimal on paper. From Markowitz's efficient frontier to Black-Litterman's view incorporation, risk parity's equal-risk philosophy, and the practical challenges of factor exposure management and rebalancing in live markets. The gap between textbook portfolio theory and what actually works is large — this module covers both.

Learning Objectives

Construct the efficient frontier and identify the tangency portfolio
Diagnose and correct the estimation error problem in mean-variance optimization
Apply the Black-Litterman model to incorporate active views into portfolio construction
Build a risk parity portfolio and understand when it outperforms MVO
Manage intended and unintended factor exposures in a multi-asset portfolio
Design a rebalancing policy that balances tracking error and transaction costs

1. Mean-Variance Optimization

The Markowitz Framework

Harry Markowitz (1952) formalized the intuition that diversification reduces risk. His insight: what matters is not the risk of each asset in isolation, but how assets' returns move together.

The optimization problem:

Minimize portfolio variance:

min   wᵀΣw
 w

subject to:
  wᵀμ = μ_target     (target return constraint)
  wᵀ1 = 1             (weights sum to 1)
  wᵢ ≥ 0 (optional)  (long-only constraint)

Where:

w = vector of portfolio weights
Σ = covariance matrix of returns
μ = vector of expected returns

The Efficient Frontier

Solving the optimization across all possible target returns traces the efficient frontier — the set of portfolios offering the highest return for each level of risk.

Expected Return
     |              * Efficient Frontier
     |           *
     |        *
     |      *
     |    *
     |  * ← Minimum Variance Portfolio
     | *
     |*
     +--------------------------- Risk (Std Dev)

Key portfolios on the frontier:

Minimum variance portfolio (MVP): Lowest risk achievable. No return target imposed.
Tangency portfolio: Where a line from the risk-free rate is tangent to the frontier — the highest Sharpe ratio portfolio.
Capital Market Line (CML): The line from the risk-free rate through the tangency portfolio. Every point on the CML is a combination of the risk-free asset and the tangency portfolio.

CAPM implication: If all investors use MVO with the same inputs, they all hold the same tangency portfolio — the market portfolio. The CML becomes the Security Market Line.

Practical MVO: Python Implementation Sketch

import numpy as np
from scipy.optimize import minimize

def portfolio_variance(weights, cov_matrix):
    return weights @ cov_matrix @ weights

def efficient_frontier(mu, cov, n_points=100):
    n = len(mu)
    results = []
    
    for target_return in np.linspace(mu.min(), mu.max(), n_points):
        constraints = [
            {'type': 'eq', 'fun': lambda w: w @ mu - target_return},
            {'type': 'eq', 'fun': lambda w: w.sum() - 1}
        ]
        bounds = [(0, 1)] * n  # long-only
        w0 = np.ones(n) / n
        
        res = minimize(portfolio_variance, w0, args=(cov,),
                      method='SLSQP', bounds=bounds,
                      constraints=constraints)
        
        results.append({
            'weights': res.x,
            'return': target_return,
            'vol': np.sqrt(res.fun)
        })
    return results

KSE-100 efficient frontier: Using 5 years of monthly KSE-100 constituent returns, the tangency portfolio would typically be concentrated in 8–15 stocks (OGDC, HBL, MCB, Engro, Lucky) — not the "diversified" portfolio the theory suggests, because Pakistani stocks are highly correlated during stress periods.

2. The Estimation Error Problem

Why MVO Breaks in Practice

Markowitz optimization requires two inputs:

Expected returns μ — extremely hard to estimate precisely
Covariance matrix Σ — more stable, but still noisy for large universes

The fundamental problem: optimization amplifies estimation errors. The optimizer treats noise in your return estimates as signal and constructs extreme positions to exploit it.

Result: MVO portfolios in practice are:

Highly concentrated (optimizer puts everything in the stocks with the highest estimated μ)
Unstable (small changes in estimates produce dramatically different weights)
Terrible out-of-sample (the estimated μ were wrong, the extreme bets lose)

Famous result (Michaud 1989): MVO portfolios frequently perform worse out-of-sample than an equally-weighted portfolio. The 1/N portfolio is a tough benchmark.

Solution 1: Ignore Returns, Optimize on Risk Only

The minimum variance portfolio requires only Σ (no return estimates). Because estimating Σ is more reliable than estimating μ, the MVP is more stable and better out-of-sample than tangency portfolio estimation.

Finding: In many studies, the minimum variance portfolio beats the market-cap weighted index out of sample, particularly in bear markets.

Solution 2: Michaud Resampling

Resampled efficiency (Michaud, 1998):

Generate N simulated return/covariance scenarios by bootstrapping from your data
Optimize on each simulation
Average the resulting weights across all N optimizations

The averaging smooths out the optimizer's tendency to exploit noise. Resampled portfolios are more stable and better diversified than single-optimization MVO.

Solution 3: Ledoit-Wolf Covariance Shrinkage

For a universe of n assets, the sample covariance matrix requires estimating n(n+1)/2 parameters. With 100 assets and 3 years of monthly data (36 observations), you have 5,050 parameters and only 36 data points — massive estimation error.

Ledoit-Wolf shrinkage (2004):

Σ_shrunk = (1 − α) × Σ_sample + α × Σ_target

Shrink the sample covariance toward a structured target (e.g., constant correlation model or single-factor model). The shrinkage coefficient α is chosen optimally to minimize expected estimation error.

Effect: Off-diagonal correlations are pulled toward the mean (reducing extreme correlation estimates), diagonal variances are more stable. Out-of-sample portfolio performance improves significantly.

Python: sklearn.covariance.LedoitWolf implements this directly.

3. Black-Litterman Model

The Problem with Standard MVO

When you run MVO with your estimated μ, the optimizer ignores the market's information. It also requires you to have a view on every asset — most investors have strong views on a few securities but no particular view on most.

Black and Litterman (1990/1992) fixed this by combining:

Prior: The market's implied returns (reverse-engineered from market cap weights)
Views: Your own active views expressed with confidence levels

Step 1: Reverse Optimization (Implied Equilibrium Returns)

Start from the assumption that the market portfolio (market-cap weights) is the "optimal" portfolio. Back out the expected returns that would make it optimal:

Π = λ × Σ × w_mkt

Where:

Π = implied equilibrium excess returns
λ = risk aversion coefficient (typically 2.5–3.5)
Σ = covariance matrix
w_mkt = market-cap weights

These implied returns are the prior — what the market collectively believes.

Step 2: Express Your Views

Views are expressed as:

P = pick matrix (which assets your view is about)
Q = view returns (what you expect)
Ω = uncertainty of views (diagonal matrix — variance of each view)

Example views:

View 1: "OGDC will outperform PPL by 3% over the next month" → P = [+1, −1, 0, 0, ...], Q = 3%
View 2: "MCB will return 8% next month" → P = [0, 0, +1, 0, ...], Q = 8%

View confidence (Ω): Low Ω = high confidence (tight distribution). High Ω = low confidence.

Step 3: Combine to Get Posterior Returns

Black-Litterman posterior expected returns:

μ_BL = [(τΣ)⁻¹ + PᵀΩ⁻¹P]⁻¹ × [(τΣ)⁻¹Π + PᵀΩ⁻¹Q]

Where τ (tau, typically 0.025–0.05) scales the uncertainty of the prior.

Result: The BL posterior blends the market's view with yours, weighted by their relative confidence. If you express no views, BL returns the implied equilibrium and the optimal portfolio is market-cap weights. If you have a strong view, the portfolio tilts toward it proportionally.

FERROQUANT BL application:

Prior: KSE-100 market-cap weights
View 1: FERROQUANT's quant model says OGDC will outperform by 4% next month (IC = 0.3 → moderate confidence)
View 2: Macro view that Pakistani banks will outperform by 5% given IMF deal progress (high confidence)
BL combines these views with market equilibrium → tilted portfolio without extreme weights

Why BL Outperforms Pure MVO

Starts from a sensible prior (market) rather than noisy return estimates
Views are expressed with explicit confidence — you can't accidentally put infinite faith in a signal
When you have no views, you hold the market — not some extreme optimizer artifact
Portfolios are more stable and intuitive than MVO outputs

4. Risk Parity & Risk Budgeting

The Problem with Market-Cap Weighting

In a traditional 60/40 portfolio (60% equities, 40% bonds):

Equity volatility: ~15% annualized
Bond volatility: ~5% annualized
Correlation: ~−0.3

Despite being 40% of the portfolio by weight, bonds contribute only ~10% of total portfolio risk. Equities dominate. The 60/40 portfolio is, in risk terms, ~90% equities.

Risk Parity: Equal Risk Contribution

Risk parity allocates so each asset contributes equally to total portfolio risk.

Marginal Risk Contribution (MRC):

MRC_i = ∂σ_p/∂w_i = (Σw)_i / σ_p

Risk Contribution (RC):

RC_i = w_i × MRC_i

Total risk = Σ RC_i = σ_p

Equal risk contribution condition:

RC_i = RC_j   for all i, j

This requires solving for weights numerically. In a 2-asset portfolio, equal risk means:

w₁σ₁ = w₂σ₂   →   w₁/w₂ = σ₂/σ₁

Higher volatility asset gets lower weight — the opposite of momentum (which buys what has gone up).

Risk Parity in Practice

The leverage problem: Risk parity portfolios are heavily overweight bonds (low volatility) and underweight equities. To achieve equity-like returns, they must use leverage (typically 2–4× on the bond allocation).

Historically: Risk parity has performed well because:

Bonds added return (falling rate environment 1980–2020)
True diversification across risk sources
Low volatility in portfolio returns

Post-2022 challenge: When bonds and equities both fall together (rising rates), risk parity's diversification benefit breaks down. The 2022 environment was extremely bad for risk parity strategies.

Bridgewater's All Weather Portfolio: The most famous risk parity fund. Ray Dalio's concept of balancing portfolio across four economic environments: growth up/down × inflation up/down.

Risk Budgeting

A generalization of risk parity: instead of requiring equal risk from each asset, set a target risk budget:

RC_i = b_i × σ_p    where Σb_i = 1

This allows you to express views through risk allocation rather than return forecasts. If you're bullish on Pakistani equities, increase their risk budget from 25% to 35%.

FERROQUANT risk budget example:

Asset Class	Risk Budget
Pakistan equities (KSE-100)	35%
Pakistan fixed income (PIBs)	20%
Gulf equities	25%
Commodities (oil, gold)	15%
Cash/T-bills	5%

5. Factor Exposure Management

Intended vs Unintended Factor Tilts

When a portfolio manager makes active bets, they often take on factor exposures they didn't intend.

Example: FERROQUANT's stock picker selects 20 Pakistani equities based on fundamental analysis. The portfolio ends up with:

Intended alpha: stock-specific insights
Unintended tilts: high beta (all picks are cyclical), value tilt (screens favor low P/B), sector concentration (overweight energy and banks)

If the portfolio underperforms, was it because the stock picks were wrong, or because the factor tilts worked against it? Without decomposing, you can't know.

Factor Neutralization

Beta hedging: Remove the market (β) exposure by shorting index futures:

Hedge ratio = Portfolio Beta × Portfolio Value / Futures Contract Value

A long/short equity portfolio with β = 0 is "market neutral" — its returns are driven purely by stock selection, not market direction.

Sector neutralization: For each sector, go long and short within the sector so net sector weight = 0. Eliminates sector rotation risk.

Factor neutralization: Construct a portfolio where exposure to each factor (value, momentum, size, quality) is zero. Returns reflect only your stock selection skill.

The trade-off: Each constraint reduces freedom and may reduce expected return. Market neutralization eliminates market return. Factor neutralization eliminates factor premia. Only do it for factors you don't want to own.

Factor Attribution

Decompose portfolio return into factor contributions:

R_portfolio = α + β_mkt × R_mkt + β_value × R_value + β_momentum × R_mom + ε

Estimate betas using regression. Then:

Factor return = β_factor × realized factor return
Alpha = residual return not explained by factors

This tells you whether outperformance came from:

Market timing (β_mkt timing)
Factor tilts (value, momentum exposure)
True alpha (stock selection)

6. Portfolio Rebalancing

Why Rebalancing Is Necessary

Over time, winners grow in weight and losers shrink. Without rebalancing:

Factor exposures drift from targets
Concentration risk increases
The portfolio no longer reflects your investment thesis

Example: A PKR 10B portfolio starts 50/50 equities/bonds. After a 30% equity rally, equities become 57% of the portfolio. Without rebalancing, you are unintentionally overweight equities — more risk than intended.

Rebalancing Triggers

Calendar rebalancing: Rebalance at fixed intervals (monthly, quarterly). Simple but ignores how much the portfolio has drifted.

Threshold (tolerance band) rebalancing: Rebalance only when an asset class drifts beyond a threshold (e.g., ±5% of target).

More efficient: only trade when necessary
Requires monitoring
Typical bands: ±3–5% for asset classes, ±1–2% for individual stocks

Optimal rebalancing (cost-benefit): Trade only when the cost of not rebalancing (tracking error relative to target) exceeds the transaction cost of rebalancing. Requires estimating both.

Best practice for FERROQUANT: Threshold rebalancing at ±3% for asset class level, ±1.5% for individual positions, reviewed daily by risk system.

Transaction Cost-Aware Rebalancing

Blind rebalancing ignores transaction costs. In practice, you should:

Calculate the current allocation and the target
Identify trades needed to reach target
Estimate transaction costs for each trade
Trade only where benefit (variance reduction, factor alignment) exceeds cost

Netting: Before executing, look for natural offsets. If buying Pakistan equities for rebalancing and simultaneously receiving redemption proceeds, net the two flows instead of executing separately.

Tax-Aware Rebalancing

For taxable investors (less relevant for institutional funds, very relevant for HNWI portfolios):

Loss harvesting: When a position is below cost, sell to crystallize the loss (reduces capital gains tax), then buy a similar (but not identical) instrument to maintain exposure
Gain deferral: Avoid selling positions with large embedded gains unless absolutely necessary
Lot optimization: When selling a partial position, choose the lot with the highest cost basis to minimize capital gains

Pakistan CGT: Capital gains tax in Pakistan depends on holding period:

< 1 year: 15%
1–2 years: 12.5%
2–3 years: 10%
3 years: 0% (exempt)

For FERROQUANT managing taxable accounts, hold periods beyond 3 years eliminate CGT entirely on equity gains.

Transition Management

When a portfolio changes manager or strategy, large-scale portfolio restructuring is needed. This is "transition management":

The existing portfolio (legacy) must be unwound
The new portfolio must be built
Both simultaneously, to minimize exposure to the "transition" portfolio (neither old nor new)

Transition risk: During transition, you may have no intended exposures — just a mix of what's being sold and bought. This unintended exposure needs to be minimized.

Transition manager role: Specialist firms manage transitions using crossing (netting buys/sells in-house), derivatives overlays, and careful sequencing to minimize cost and risk.

Self-Assessment

FERROQUANT runs a PKR 5B long-only Pakistan equity fund. The team estimates the following for 5 liquid KSE stocks:

Stock Expected Return (ann.) Volatility (ann.)
OGDC 14% 22%
HBL 18% 28%
Lucky Cement 22% 35%
Engro 16% 25%
PSO 12% 30%

Average pairwise correlation: 0.55. Risk-free rate: 13%.

(a) Without running the full MVO, explain why MVO might put excessive weight on Lucky Cement. What is the fundamental problem with trusting this result? (b) Why would a minimum variance portfolio likely have lower HBL and Lucky Cement weights despite their high expected returns? (c) If Ledoit-Wolf shrinkage is applied to the correlation matrix, what would happen to the 0.55 average pairwise correlation and why? (d) An equally-weighted portfolio has Sharpe ratio 0.28 in out-of-sample testing. The MVO tangency portfolio has Sharpe ratio 0.18 out-of-sample. What does this tell you about estimation error in the expected return inputs?
FERROQUANT wants to apply Black-Litterman to the same 5-stock universe. Market-cap implied returns are approximately equal to current expected returns above. FERROQUANT has two views:
- View A: "HBL will outperform OGDC by 6% over the next year" (medium confidence — IC ≈ 0.25)
- View B: "Lucky Cement will return 28% next year driven by CPEC infrastructure" (high confidence — IC ≈ 0.50)
(a) Express View A as a P matrix row (5-element vector) and a Q value. (b) How should you set Ω (uncertainty) differently for View A vs View B? (c) Without exact calculation, describe qualitatively what the BL posterior return for Lucky Cement would look like relative to the implied equilibrium return of 22%. (d) If FERROQUANT has no views at all, what does BL produce as the optimal portfolio, and what is the intuition?
FERROQUANT manages a PKR 3B risk parity portfolio across 4 asset classes. Current volatilities and correlations:

Asset Volatility Target Weight (market cap)
Pakistan Equities 20% 40%
Pakistan Bonds (PIBs) 6% 30%
Gulf Equities 18% 20%
Commodities (oil/gold) 22% 10%

(a) Calculate the approximate risk contribution of each asset class under market-cap weights (simplify: assume zero correlation between asset classes). (b) Which asset class dominates risk in the market-cap weighted portfolio? (c) Solve for the equal risk contribution weights (zero correlation assumption). Show your work. (d) The risk parity portfolio requires 2.5× leverage on Pakistan bonds to achieve a 12% annualized return target. Explain the rationale and identify the key risk this leverage introduces.

Asset	Volatility	Target Weight (market cap)
Pakistan Equities	20%	40%
Pakistan Bonds (PIBs)	6%	30%
Gulf Equities	18%	20%
Commodities (oil/gold)	22%	10%

Stock	Expected Return (ann.)	Volatility (ann.)
OGDC	14%	22%
HBL	18%	28%
Lucky Cement	22%	35%
Engro	16%	25%
PSO	12%	30%