How we set the default return forecasts.
Every projection in Portfolio Lab rests on an expected return for each asset class. This page explains where those numbers come from, why the default is an average rather than one firm’s book, and what we found when we tested whether that choice was right.
The problem this solves
Nobody knows what markets will return. The professional answer is a set of capital market assumptions: a firm publishes its long-run expected return for each asset class, usually once a year, and everything downstream is built on those figures.
Two things go wrong with using one firm’s set. The first is that the firms disagree, sometimes by more than the return they are forecasting. The second is that a forecast is struck on a particular day against that day’s prices, and it is quoted for a year afterwards as though the prices had not moved.
Whose forecasts we carry
4 houses publish enough detail to be usable, and none of them covers everything:
- J.P. Morgan LTCMA publishes 42 of 42 asset classes
- BlackRock CMA publishes 20 of 42 asset classes
- Research Affiliates AAI publishes 40 of 42 asset classes
- AQR CMA publishes 17 of 42 asset classes
We also cite GMO’s 7-year forecast on the comparison page but do not run on it. They publish eight of our asset classes, which is too few to build a portfolio from: choosing GMO would mean running mostly on other firms’ numbers under GMO’s name.
Why the default is the average
Committing to a single house is a stronger claim than it looks. It says you trust one firm’s emerging market debt call because you trust their US equity call, and nobody actually believes that. So each asset takes the mean of the houses that publish a view on it.
Averaging only the houses that have a view
This restriction is the part that does the work. Coverage runs from 17 to 42 asset classes, so a naive mean over 4 sources would report an asset that only one firm forecasts as 4 houses agreeing. That is one opinion counted 4 times, wearing the authority of a consensus. Where a single house is the only voice, the app says so on the row rather than calling it an average.
Averaging the numbers, not the portfolios
The average is taken across expected returns, and separately we tested averaging the portfolios each firm’s numbers produce. Both work; the second is closer to what the academic literature calls resampled efficiency. The figures below are from the first.
Bringing each forecast up to today’s prices
Averaging happens after marking, never before. Each house struck its numbers on its own date, so a mean of raw figures would be an average across time as much as across opinion, and there would be no single date to mark the result from afterwards.
Marking means different things for different assets, and the difference matters:
Government bonds, inflation-linked bonds and cash
Observable outright. Over a horizon near the index’s own duration, the starting yield predicts the return closely enough that no forecasting happens at all. These are replaced with today’s figure. Cash currently sits at 4.19 percent, against the 3.10 percent in the registry’s own base table.
Corporate and emerging market credit
A credit return decomposes into three terms: the risk-free rate, the option-adjusted spread, and the expected credit loss. FRED publishes the first two daily. Only the loss is a view, it is the smallest of the three, and it moves over decades rather than months.
So rather than imposing our own loss assumption, we back out each house’s. Their published return against the yield they were looking at on their own strike date gives their view of default risk, and that view is kept and applied to today’s yield. The house’s judgement survives and only the market moves. Two houses that struck on the same day and disagree about credit still disagree by exactly the same amount after marking.
Equities
A long-horizon equity forecast travels from today’s multiple to an assumed terminal multiple, and only the starting point goes stale. Refreshing it needs no view about where multiples end up, because the terminal assumption cancels out of the difference. The change comes from the index’s own price move since the strike date, net of earnings growth.
Assets with no price of their own
Listed property, infrastructure, private equity, hedge funds and local currency emerging market debt cannot be marked directly. A private mark is appraised rather than traded; a hedge fund composite is a fee structure over other people’s positions. But leaving them at last year’s figure while everything around them moves is not neutral either. It quietly reprices them against the rest of the table, and an optimiser acts on exactly that.
So they move by what they are measurably made of. Each has a regression against equities and duration, and where that model explains at least 30 percent of the asset’s movement, the asset is marked by its replicating portfolio: its equity beta applied to the equity mark, funded at its own measured duration.
39 of 42 asset classes are refreshed. Prices were last read on 2026-08-23.
What our own tests showed
Marking removes disagreement that was never real
After marking, US Treasuries, cash and world government bonds show a spread of exactly zero across all 4 houses. They were never really disagreeing about those. They were quoting different Tuesdays. Genuine disagreement survives where it should: private equity still spans 10.3 percentage points across 4 houses, and US large cap 3.8 points across 4.
Averaging beats committing to any single house
We solved the same portfolio 4 times, once on each house’s numbers, then scored every candidate answer under every house’s assumptions. The measure is maximum regret: how far short of the best available answer you fall if the house you did not pick turns out to be right.
Averaging the 4 portfolios carried a maximum regret of 0.022 Sharpe. The best single house carried 0.023 and the worst 0.052. Equal weighting across all assets, which ignores the forecasts entirely, carried 0.099. So averaging beat committing to any one firm without needing to know which firm was right, and it held nearly twice as many effective positions as the most concentrated single-house answer.
The weights move much more than the answer
On one representative run the 4 houses disagreed by up to 54 percentage points on a single holding, and the portfolios that came out of that disagreement landed within 0.44 percentage points of expected return a year of each other. Mean-variance objectives are flat near their optimum, so weights slide a long way while the portfolio on the end of them barely moves. The instability is real and it is nearly free.
What the research says
The choice to average is not ours. Forecast combination is one of the more durable findings in the forecasting literature: Timmermann’s survey in the Handbook of Economic Forecasting (2006) documents that simple equal-weighted combinations repeatedly outperform sophisticated schemes that try to weight forecasters optimally, because those weights are themselves estimated with error.
Applied to this exact problem, a study of 19 firms’ 2013 ten-year forecasts against what actually happened found that averaging prevented severe misses, including a roughly three point over-prediction on emerging market equities. The same work is candid that in most cases the average was still an unreliable guide to the realised return, which is a caveat we repeat rather than bury.
Mike Sebastian’s work on the accuracy of capital market assumptions finds forecast errors ranging from 0.1 to 3 percentage points across asset classes, and singles out private equity and hedge funds as the categories where assumptions are least reliable, because manager dispersion is wide and there is no investable market to anchor to. Our own spread figures agree: private equity is the widest disagreement in the table at 10.3 points.
Sebastian’s practical advice is to de-emphasise capital market assumptions in favour of starting from the market portfolio and adjusting for genuine views, and to be aware of how the assumptions you use compare with industry averages. The second half of that is what the comparison tool exists for.
What you can change
The average is a default, not a claim that averages are correct. In the app’s Assumptions tab you can see all 4 forecasts side by side for every asset class, run the whole tool on any single firm, or type your own numbers over the top. Whichever you choose is stamped on every projection and on the client report, so a figure never carries a firm’s name unless that firm actually published it.
Known limitations
- The 3 unmarked assets are compared against 39 that have been refreshed. That asymmetry favours whichever side the market has moved against.
- The emerging market spread we mark against is a corporate index, which is investment grade heavy, so it understates a sovereign index.
- Averaging across houses with different horizons mixes a 10-year forecast with a 10 to 15-year one. We note the horizons rather than adjusting for them.
- An average of four is still four opinions. Every caveat that applies to a capital market assumption applies to a mean of them.