Why We Average Capital Market Assumptions Instead of Picking One

23 August 2026 · 7 min read

Ask four investment firms what private equity will return over the next decade and you get 4 answers spanning 10.3 percentage points. BlackRock says 13.5 percent a year. Research Affiliates says 3.2 percent. Same asset class, same decade, four times the return.

These published forecasts are called capital market assumptions, and almost every financial plan you will ever be shown rests on one firm’s set of them. Which raises an uncomfortable question that the firms themselves never answer: if they disagree this much, whose numbers should you actually use?

We spent a while testing that, because Portfolio Lab has to pick a default and we did not want to pick one by taste. This is what we found.

The short answer

That is now the default in Portfolio Lab. Every asset class takes the mean of the houses that publish a view on it. You can still switch the whole tool to any single firm, or type your own numbers, and the app tells you which you are using.

Some of the disagreement is not real

Before deciding how to combine the forecasts, we had to separate the genuine disagreement from the fake kind. A firm publishes its assumptions once a year, on a particular day, against that day’s bond yields and share prices. Twelve months later the figure is still being quoted as though nothing had moved.

So we bring each firm’s numbers up to today’s prices before comparing them. Government bonds are the clean case: what a bond fund pays you today is the best available guide to what it will return, so there is no forecasting involved at all. Once you do that, all 4 houses agree on US Treasuries to within 0.00 percentage points.

What survives that correction is real. US large cap still spans 3.8 points across 4 houses, because the firms genuinely disagree about whether today’s valuations matter. Private equity spans 10.3, gold spans 5.3, and no amount of arithmetic will reconcile those.

Why averaging wins

We ran the same portfolio problem 4 times, once on each firm’s numbers, which gave 4 different answers. Then we scored every answer under every firm’s assumptions and asked a simple question: if you commit to one firm and a different one turns out to be right, how much worse off are you?

Averaging the four portfolios came out ahead of committing to any single firm. The gap over the best single choice was small, which is the point: the average was never the worst answer, and picking a firm in advance risks being. Ignoring the forecasts entirely and splitting evenly across assets was substantially worse than all of them.

The averaged portfolio was also better diversified. The most concentrated single-firm answer collapsed into fewer than three meaningful positions; the average held nearly five.

This is a known result, not our idea

Combining forecasts is one of the more durable findings in forecasting research, well outside finance. Allan Timmermann’s survey in the Handbook of Economic Forecasting documents that simple equal-weighted averages repeatedly outperform sophisticated schemes that try to weight forecasters by past skill. The reason is a little bleak: those weights have to be estimated too, and the estimation error usually costs more than the cleverness gains.

Applied to this exact problem, a study comparing 19 firms’ 2013 ten-year forecasts against what actually happened found that averaging prevented the severe misses, including an over-prediction on emerging market equities of around three points a year. It is also honest that in most cases the average was still an unreliable guide to the realised return, which is worth repeating rather than burying.

Mike Sebastian’s research on the accuracy of published assumptions found forecast errors between 0.1 and 3 percentage points a year, and singled out private equity and hedge funds as the least reliable categories because manager dispersion is wide and there is no investable index to anchor to. Our own figures agree with him: private equity is the widest disagreement in our table at 10.3 points.

The trap in averaging

One detail turned out to matter more than anything else in the arithmetic. The four firms do not cover the same ground:

• J.P. Morgan publishes 42 of 42 asset classes
• Research Affiliates publishes 40 of 42 asset classes
• BlackRock publishes 20 of 42 asset classes
• AQR publishes 17 of 42 asset classes

If you average blindly across four sources, an asset that only one firm forecasts comes out looking like four firms in perfect agreement. It is one opinion counted four times, wearing the authority of a consensus, and it is the most confident-looking number on the page for exactly the wrong reason.

So the average only ever includes firms that actually published a view on that specific asset. Where a single house is the only voice, we say so on the row instead of calling it a consensus.

How much does any of this change?

Less than the size of the disagreement suggests, which surprised us. On one run the four houses produced weightings that differed by up to 54 percentage points on a single holding, and the portfolios that came out of those wildly different weightings landed within 0.44 percentage points of expected return a year of each other.

The reason is a well-known property of portfolio optimisation: near the best answer the objective is almost flat, so the weights can slide a long way while the portfolio on the end of them barely changes. Corporate bonds, high yield and emerging market debt are close enough substitutes that swapping between them costs almost nothing.

What to do with this

If you are building a plan and someone hands you a projection, the useful questions are whose assumptions it rests on, when those assumptions were struck, and how far the answer moves if you use somebody else’s. A plan that only works on one firm’s numbers is telling you something.

You can check all of this yourself. The capital market assumptions comparison shows every firm’s figure side by side for all 42 asset classes, and the full methodology sets out exactly how the default is built, including what we cannot correct and why.