Portfolio Planning

Problem 1: No reliable return data for alternatives

Traditional portfolio optimization requires return time series. For public equities and bonds, these exist. For private equity, venture capital, and real estate funds, they do not. Reported NAVs are smoothed and lag reality. Listed PE indices cover only a fraction of the market.

The result: most SAA tools either exclude alternatives entirely, or include them as a single bucket with assumed returns and artificial correlations. Neither is acceptable for portfolios where alternatives represent 20-50% of assets.

Our approach: replicating portfolios

We construct synthetic return time series for each alternative asset class by building portfolios of liquid assets (Capital Gain Strips) that replicate the actual cash flow profile of illiquid funds. The methodology follows Gupta and Van Nieuwerburgh (2021, Journal of Finance).

For each of 10 alternative asset classes (Buyout Large, Mid, Small, Growth, VC Early, VC Late, Private Debt, Real Estate, Infrastructure, Secondaries) and each quarter, we identify the combination of public risk factors that best replicates the actual fund cash flows. The empirical basis: 5,000+ funds with 30+ years of net cash flow data .

The result: continuous, comparable return time series for every asset class, public and private. For the first time, we can compute real correlations between alternatives and traditional assets, not the artificial low correlations that come from smoothed NAV reporting.

Three dimensions, not two

Standard portfolio optimization uses two dimensions: return and risk (typically volatility). This works for liquid assets where you can exit at any time. For portfolios with significant alternative allocations, a third dimension is essential: liquidity .

The efficient frontier is no longer a curve in two dimensions. It becomes a surface in three dimensions : for each combination of risk tolerance and liquidity tolerance, there is an optimal portfolio. Investors who can accept more illiquidity unlock higher-return allocations.

Problem 2: Cash flow forecasting ignores market conditions

Once you have a target allocation, you need to plan commitments and manage liquidity. This requires forecasting when capital gets called and when it comes back. Existing models have clear limitations:

The Yale model (Takahashi-Alexander, 2002) is widely used but deterministic. It cannot adapt to changing market conditions. Stochastic models add randomness but ignore external factors. Both treat each fund in isolation.

Our approach: ML with macro variables

Our machine learning model integrates fund characteristics with macro-economic and market variables that actually drive PE cash flow timing:

Market conditions (VIX, CFNAI), public equity dynamics (S&P 500, NASDAQ, Fama-French factors), and credit conditions (yield curve spread, risk-free rate). The model learns how these factors influence capital calls and distributions at different stages of a fund's lifecycle.

The result: up to 10% lower prediction error during critical fund years (4-6) and up to 25% lower downside risk compared to the next best model. This means fewer surprises when capital calls arrive.

Why macro variables matter

Cash flow behavior is not constant. In the early fund years (1-3), broad market conditions are the strongest predictor of capital call timing. From year 3 onwards, a fund's own cash flow history becomes dominant. In the harvest period (year 7+), exit activity is heavily influenced by public equity valuations and credit conditions.

A model that ignores these dynamics will systematically mispredict cash flows during exactly the periods when accuracy matters most: market stress, credit tightening, or rapid exits in hot markets.

Three decisions. One platform.

Everything on the platform is designed to inform three concrete decisions:

Academic foundations