Research lab · Asset allocation
A mean-variance workbench for nine asset classes and seven model profiles. Build expected returns from building blocks, set volatilities and correlations, and read the return, risk and wealth range each mix implies.
Bear and bull rescale each asset's expected return in excess of cash. At 70% an asset keeps 70% of its premium, at 130% it earns 130% of it. Cash itself is unchanged.
Select a profile with the buttons above. Weights are editable. A column that does not add to 100 is rescaled to 100 in the calculations.
Annual, nominal, in percent. Bond and equity expected returns are the sum of the building blocks below. Changing the cash rate shifts the risk-free block of every bond class by the same amount.
VaR and expected shortfall are losses on the initial amount over one year, net of the expected return. A negative value means the 5% tail still ends above the start.
The median path compounds at the expected return less the drag from volatility, so it sits below a path compounded at the arithmetic mean. The percentile range comes from volatility alone, with the expected return held at the base case.
The frontier is the highest expected return available at each level of volatility with no shorting and no leverage. It moves a lot when the inputs move a little, so read it as a consistency check on the assumptions, not as a target.
Risk share is each asset's part of portfolio variance. It adds to 100% and can be negative for an asset that hedges the rest.
Expected returns. Bonds and equities are built from additive blocks. Alternatives and cash are entered directly.
bond return = risk-free yield + spread + roll-down + valuation (rates) + valuation (spread) + default loss
equity return = dividend yield + real earnings growth + inflation + valuation change + currency
Portfolio. With weights w (rescaled to sum to one), expected returns μ, volatilities σ and correlations ρ:
μp = Σi wi μi
σp2 = Σi Σj wi wj σi σj ρij
Sharpe = (μp − rcash) / σp
risk sharei = wi (Σw)i / σp2
Scenarios. Bear and bull change expected returns only. Volatilities and correlations stay put. With multiplier m (0.70 for a 70% bear case):
μi(m) = μi + |μi − rcash| · (m − 1)
The absolute value keeps the bear case below the base case for an asset whose premium over cash is negative.
Wealth and tail risk. The portfolio is held at constant weights, which means it is rebalanced back to target every period at no cost. Its annual gross return is taken as lognormal with the same mean and variance as above, independent from year to year:
s2 = ln(1 + σp2 / (1 + μp)2), g = ln(1 + μp) − s2/2
median wealtht = W0 · exp(g t)
percentile q of wealtht = W0 · exp(g t + zq s √t)
VaR95 = 1 − exp(g − 1.645 s)
ES95 = 1 − (1 + μp) · Φ(−1.645 − s) / 0.05
P(wealthT < W0) = Φ(−g √T / s)
Drawdown. 2,000 simulated paths of monthly log returns with mean g/12 and standard deviation s/√12, from a fixed random seed so the same inputs always give the same answer. The maximum peak-to-trough loss is recorded on each path, and the median and 95th percentile across paths are reported.
Frontier. For a range of risk tolerances λ, the weights solve the problem below. The correlation matrix must be positive semi-definite for any of this to be meaningful. The page checks its smallest eigenvalue and can repair it by lifting every eigenvalue below a small positive floor of 0.02 up to that floor, rescaling the diagonal back to one and rounding to two decimals. The floor is positive, not zero, so the repaired matrix stays invertible. If the rounding pushes an eigenvalue back to zero or below, the floor is raised in steps of 0.02 and the repair repeated.
minimise ½ w′Σw − λ μ′w subject to wi ≥ 0, Σi wi = 1
Limits. Returns are not lognormal and not independent across years: real tails are fatter, volatility clusters and correlations rise in stress, so the tail and drawdown figures here are a floor more than an estimate. A single period model has no transaction costs, taxes, fees or liquidity limits. The two custom profiles are fixed example mixes, not the output of an optimiser. The building blocks treat valuation change as an annual average over the horizon, which hides its timing.
Educational tool. All inputs are illustrative assumptions, not live data and not a house view. Not investment advice.