Assumptions & the market model

Every plan needs an opinion about the future — what a stock market returns, how bumpy that ride is, and how fast prices rise generally. The plan's Assumptions & market model section is where that opinion lives: a return and a volatility for six asset classes, plus inflation, and every one of them moves the projected success percentage. The card offers two named sets of them; this page sets both beside dated, published figures — long-run history and current forecasts — so you can judge them and choose your own. One field there is not an opinion at all — what you pay in fees — and it has a section of its own below.

Every outside figure below carries its source, its date, and two labels that matter more than they look: real or nominal (after or before inflation) and compound or arithmetic (explained under Which "return" to type). Comparing two numbers that differ on either label is the most common way to get this wrong. None of this is advice; it is what the numbers are and where they sit.

The defaults

The workbench carries two named sets, and the Assumptions & market model card prints both beside the fields. Each cell is expected return · volatility.

A new plan starts on Conservative. Set to Conservative and Set to Consensus, at the top of the card, each put all six rows and inflation on that set in one click; the card says which set the plan is on, or Custom once a value differs from both, with each changed field marked. Nothing is saved until you press Save plan, and Discard puts the old values back.

Plus three rates that apply across the whole plan rather than to one asset class:

• Inflation — 2.5% a year, in both sets.
• Annual fees — what you pay, and 0% until you enter it. It belongs to neither set: see Fees are yours to enter.
• Medical inflation — 1.5 percentage points above general inflation, compounding only the health insurance line.

The six asset-class rows, plus inflation and fees, are edited in the plan's Assumptions & market model section. Medical inflation isn't there, even though it's also a rate: it's edited in Health insurance, beside the premiums it multiplies (see Households and the plan). Nothing here is fixed — the workbench exists so you can try your own view instead of either of these.

Where the two sets come from

Consensus, the set to compare against, is the industry's own average: the 20-year arithmetic expected return and the standard deviation for each asset class from the 2026 edition of Horizon Actuarial's Survey of Capital Market Assumptions (Exhibit 17 — 43 investment advisors, of whom 29 gave the 20-year figures used here), rounded to the nearest 0.25 for returns and 0.5 for volatility. The Real estate & alt row is a catch-all, so it takes the midpoint of the survey's real-estate and commodities lines.

Conservative, the set a new plan starts on, sets each return 1 to 1.75 points under that, and never a volatility more than 2 points calmer. It is the cautious first-decade view: the one thing a single return per asset class genuinely cannot say is that the first decade may not look like the average of four. Ten-year forecasts built from today's valuations run well below the 20-year figures — the evidence table below shows the gap — and the first decade of a retirement is the one that decides it. So a new plan starts on the cautious first-decade view, run as if that decade lasts; press Set to Consensus and compare to see what the long-run average would add.

What the margin is not is a charge for fat tails or for asset classes falling together in a crash. The engine already prices both: it blends volatilities in a straight line, which is what you'd get if every class fell on the same days, and then multiplies the result by 1.05 (see What the model leaves out). Cutting returns for the same two risks would charge them twice. Nor does it stand in for fees, which are now a field of their own and apply on either set.

Crypto is identical in both sets: no survey or capital-market report estimates it, so there is no consensus to set it to. Inflation is the same too — the survey's 2.44% rounds to 2.5%.

Neither set has been vetted for your household, and the most useful thing they do is disagree: snapshot a scenario on one, switch to the other, and compare the two success rates (see Monte Carlo and scenarios).

Fees are yours to enter

The figures in both tables above are market returns. What you keep is what the market returned minus what you pay to hold it — fund expense ratios, plus an advisory fee if someone manages the money for you. That is a fact about your household rather than a view about markets: it is the same in every simulated future, and you can usually look it up to the basis point.

So it is an input. Annual fees on the Assumptions card takes your all-in rate as a percentage of assets — anything from 0% to 3% — and the simulation subtracts it from the blended return in the same step that takes inflation off — it costs return, not volatility, and it costs it in every year of the plan. Left at 0 the plan charges you nothing, which is right for very few households: an index-fund portfolio held directly usually runs somewhere near 0.05–0.2%, and an advised one near 1%. The gap between those two is worth more to a long plan than most of the arguments on this page, and the field exists so that it is visible rather than buried inside a return assumption.

Two things it does not do. It doesn't change either named set — pressing Set to Consensus or Set to Conservative leaves your figure exactly where it was — and it isn't a substitute for the Conservative set, which is about the market rather than about you.

What the market model is

The thirteen numbers, and the fee rate beside them, become one simulation in five steps:

1. The six (return, volatility) pairs are blended by what you hold today (see below) into one return and one volatility. A household with nothing entered yet has no mix to blend by, so it is simulated on the Cash row alone — 2.5% / 1% at the defaults — until it adds an account.
2. Inflation comes off the return once (next section), and so do your fees, so everything after runs in today's dollars and is net of what you pay.
3. The blended volatility is multiplied by 1.05 (see What each knob does).
4. Each month of each simulated future draws its own return, independently, from a bell curve with that mean and that spread.
5. That is repeated for 1,000 futures, and the share that never run out of money is your projected success.

How the plan is simulated works the arithmetic through with numbers. The line on the card that reads "Blended … nominal · … real · volatility …" is the output of steps 1–3 for your household: the blended return before and after inflation, and the volatility after the 1.05 — the figure actually simulated, which is why it is a little higher than a by-hand blend of the Volatility column.

Nominal, then real

The figures above are nominal — the return before inflation eats into it, which is how a market return is normally quoted. The simulation doesn't use them directly, though: it takes inflation back off once, and runs the whole plan in today's dollars from there. With the defaults, that means US stocks are actually simulated at a 4.5% real return — 7% nominal, minus 2.5% inflation — not the 7% on the label.

Your annual fees come off in the same step, and the "Blended … real" line on the card is the figure after both. At 0.2% of assets that 4.5% becomes 4.3%, and it stays lower in every year of the plan.

The subtraction is a simplification. The exact conversion divides — (1 + 7%) ÷ (1 + 2.5%) − 1 = 4.39% — so straight subtraction flatters every return by about a tenth of a point at these values, and by more when inflation is set higher. It errs in the optimistic direction, and it's small next to the next section's gap.

Which "return" to type

The field wants an arithmetic average return: the plain mean of one-year returns. The return an investor actually ends up with over decades is the compound (also called geometric, or annualized) return, and it is always lower, because a 20% loss needs a 25% gain to undo. The gap is close to half the variance:

compound return ≈ arithmetic return − volatility² ÷ 2

Michael Kitces' "Volatility Drag: How Variance Drains Investment Returns" (2017) derives it, and says which one a simulation wants: "When it comes to Monte Carlo assumptions, the proper return assumption is the arithmetic mean, not the geometric average return" — the simulation generates the bumps itself, so feeding it a compound figure counts the drag twice.

On the Conservative set — the defaults — using the volatility actually simulated (the typed volatility × 1.05):

This was checked against the engine rather than taken on trust: a portfolio held entirely in US stocks, with nothing going in or out, has a median path that compounds about 1.3 points a year slower at the default volatility than at a volatility of zero — the rule of thumb's 1.2, plus a little from compounding and sampling. A test holds the engine to that.

Two things follow.

Almost every published figure is a compound one, and has to be converted before it's typed. "US stocks returned 6.6% real since 1900" is compound; so are the headline numbers in most forecast reports. To use one, add the drag back: about 1.5 points for stocks at 16–17% volatility, about 0.2 for bonds, nothing for cash. Typing a compound figure straight in understates the plan by the drag column. The reports that print both make the gap easy to see: J.P. Morgan's 2026 assumptions list US large-cap stocks at 6.70% compound and 7.94% arithmetic, at 16.47% volatility; the Horizon survey has 7.02% and 8.32%, at 16.43%. It is also why Consensus is built from the survey's arithmetic column, not the compound one its headlines quote.

The Crypto row's typical path loses money. At 60% volatility the drag is larger than the 15% return. The average of many simulated crypto paths does grow 15% a year, carried by a few enormous ones; the median path shrinks. That isn't a bug in the row, it is what that much volatility means: in a plan that holds a lot of crypto, the row acts as a drag on the typical outcome and a widening of the range, not as a growth engine.

Volatility costs a plan twice, and this is only the first: once money is being withdrawn, the order of returns matters as well — What each knob does, below.

Each default against the evidence

Compound and nominal unless marked. "History" is the UBS Global Investment Returns Yearbook 2026 (Dimson, Marsh and Staunton; US markets, 1900–2025). "JPM" is J.P. Morgan Asset Management's 2026 Long-Term Capital Market Assumptions (published October 2025, a 10–15 year outlook). "Horizon" is the survey Consensus is rounded from, 20-year horizon. "Vanguard" is the Vanguard Capital Markets Model's 10-year outlook, run of June 30, 2026. Where a report prints an arithmetic return it is given as arith. — that, not the compound figure, is the like-for-like comparison with the two sets' columns.

Where the sets sit:

• US stocks. The old shorthand is that US stocks return "about 7% real". The Yearbook's figure is 6.6%, and it is compound — so the fair comparison with the default isn't 4.5% against 6.6%. It is roughly 3.3% (4.5% less the 1.2-point drag) against 6.6%: about half of what the last century delivered, before any fee you pay. Consensus, by the same arithmetic, expects about two-thirds of it. Forecasts built from today's valuations run well below history, and both sets side with the forecasts: Conservative's 5.8% compound equivalent sits inside Vanguard's 4.2–6.2% range, and Consensus's 6.8% compound equivalent sits just above Vanguard's range and level with the compound figures J.P. Morgan (6.7%) and the survey (7.0%) publish — like against like. Consensus's 8.25% is the survey's arithmetic 8.32%, rounded; that is a different basis, and comparing the two is the mistake this page keeps warning about.
• US volatility. Consensus's 16.5% — simulated at 17.32% — is the 16.4–16.5% J.P. Morgan and the survey both use, plus the engine's 1.05. Conservative's 15% sits a little under them, inside the posture's 2-point allowance; a lower volatility is the optimistic direction, which is why the posture caps how far below it may go.
• Intl stocks. Consensus puts international half a point above the US; Conservative holds the two at parity. History and the forecasts pull in opposite directions here: the US was the twentieth century's winner, at 9.8% a year against 8.5% for developed markets as a whole (which include it), while Vanguard, J.P. Morgan and the survey all put developed ex-US above the US today, on valuations. Consensus follows the forecasts; Conservative's parity declines to bet either way. The same history is the caveat on using the US record as a guide at all: it is the record of the market that happened to win.
• Bonds. On the arithmetic basis the field uses, Consensus's 5.25% is the survey's own 5.3%, rounded, and sits a little above J.P. Morgan's 4.9%; Conservative's 4% is below every figure in the row, by 0.6 to 1.3 points. The volatility is in line in both sets.
• Cash. Consensus's 3.5% is the survey's, made with short rates above inflation — about a point of real return, where history says 0.5%. (At a volatility this low the arithmetic and compound figures are the same number, so that comparison needs no conversion.) Conservative's 2.5% equals the default inflation rate, which simulates cash as earning nothing in real terms.
• Real estate & alt. The row is a catch-all, so which comparison applies depends on what you actually hold in it: a rental property, a REIT fund, gold and a commodities fund are four different rows in any of these reports. Consensus's 7.25% is the midpoint of the survey's real-estate and commodities lines; Conservative's return is below all of them on the arithmetic basis the field uses. Both sets' volatility is that of something traded; core real estate's lower figure is appraisal-based, and smoothed.
• Crypto. There is no evidence base: no report above estimates it, and its whole history is one era. What can be measured is how volatile it has been — Fidelity Digital Assets' "A Closer Look at Bitcoin's Volatility" (May 2024) found bitcoin "three to nearly four times as volatile as various equity indices" from 2020 to 2024, with 90-day realized volatility averaging 46% over the two years before — which is the neighbourhood of both sets' 60%. The return is a placeholder, and at that volatility it isn't what drives the typical outcome anyway.

Forecasters disagree with each other by more than the two sets do: the table's compound forecasts for the same US stock market run from 4.2% to 7.0%. No number on this page is known; each is a position.

Inflation

The default is 2.5%. Beside it:

• The Federal Reserve's stated goal is 2%, measured on the PCE price index (Statement on Longer-Run Goals, reaffirmed January 2026).
• US consumer prices rose 2.9% a year from 1900 to 2025 (the Yearbook).
• The bond market's implied 10-year rate — the 10-year breakeven — was 2.33% on September 17, 2026.
• J.P. Morgan assumes 2.5%; the Horizon survey average is 2.44%.

So 2.5% is where the forecasters are, and a little below the century. What inflation does in the plan is mostly about the gap between it and your returns, with two exceptions — see What each knob does.

Medical inflation's 1.5 points over general inflation is a cautious figure rather than a measured one. The Peterson-KFF Health System Tracker (August 2024) has medical care prices up 121.3% since 2000 against 86.1% for all consumer prices, which works out to about 0.7 of a point a year. One ageing household's insurance premiums are not the medical-care price index, so that is a reference point rather than a verdict; the field is in Health insurance if you want a different number.

What each knob does

Moving an assumption changes the simulation in a specific, predictable direction:

• Expected return moves the median path and the success rate the way you'd expect — up with a higher return, down with a lower one — and the effect compounds over however many years you're simulating, so it matters more the longer your plan runs (see Time and the plan).
• Volatility does two separate things. It widens the fan chart's bands — more spread between a lucky path and an unlucky one — and it also pulls the median down: that is the drag measured above, and it needs no withdrawals at all. A plan that's actually drawing money out pays a second, larger cost from the same volatility: withdrawing while the market is down locks in a loss that a portfolio still being added to would have ridden out. That's sequence-of-returns risk, a different mechanism, and Monte Carlo and scenarios walks through a worked example of why the order returns arrive in changes the outcome even when the average is identical. Both effects lower the success rate at the same expected return — it isn't a quirk of this engine.
• Inflation mostly matters as the gap between it and your returns — raising it without touching any return is close to cutting every asset class's return by the same amount, because that gap is what the engine actually simulates against. It isn't the whole story, though: a household with a pension that has no cost-of-living adjustment (see Money in, money out) feels a second effect — a higher inflation rate erodes that check's real value faster too, on top of shrinking the return gap. And not everything in the plan is flat in today's dollars to begin with: health insurance grows in real terms and a no-COLA pension shrinks in them, so inflation is also the rate both of those move against.
• Annual fees come off the blended return alongside inflation, so a point of fees is a point of return, every year, on every path. Unlike a return assumption it isn't a guess: raising it doesn't make the plan more cautious, it makes it more accurate.
• Medical inflation compounds the health insurance line on top of general inflation, and only that line. It has no effect on spending, goals, income, or anything else in the plan.

One correction is already baked into the simulation and isn't a field you can see: the engine multiplies your portfolio's blended volatility by 1.05 before simulating. The blend itself — each class's volatility, weighted by your allocation, added straight up — is actually the volatility you'd get if every asset class moved in perfect lockstep, correlation of 1, the maximum case, rather than an assumption that they move independently: real asset classes usually aren't that correlated day to day, so on an ordinary year this blend already leans conservative. What it can't see is what happens on a bad one. There's no real covariance matrix behind it, each month is drawn from a plain bell curve with no fat tails, and it has no way to represent how normally-diversifying asset classes tend to move together anyway in exactly the crashes that hurt a retirement most. The 1.05 uplift is a flat, modest cushion against those blind spots — fat tails and crash-time correlation spikes — not a precise correction for either one. See What the model leaves out for how this sits alongside the model's other market-side simplifications.

It's blended by what you actually hold

The simulation doesn't treat these six numbers as separate return streams — it blends them by your household's current allocation across US stocks, Intl stocks, Bonds, Cash, Crypto and Real estate & alt, the same mix the portfolio dashboard shows. Two households running identical assumptions with different portfolios get different plans, because the blend itself is different. Change your allocation and the plan updates without you touching a single assumption.

Choosing your own numbers

• Pick one source and one basis, and hold to them. All history or all forecasts; all nominal, because the fields are; all converted to arithmetic, per Which "return" to type. A forecast for stocks beside a historical figure for bonds is the usual mistake — it builds a gap between the two that neither source believes in.
• A 10-year forecast is not a 40-year assumption. The reports above are decade outlooks, driven by where valuations and yields start. A plan that runs to 95 outlives them several times over, and it is reasonable to think the later decades look more like history. The engine, though, has one return per asset class for every year of the plan (see What the model leaves out), so what you are choosing is a single compromise figure. The survey's 20-year column is one published attempt at exactly that compromise, which is why Consensus uses it rather than the 10-year one.
• Sensitivity beats precision. Nobody knows whether US stocks will return 6% or 8%. What you can know is how much your plan cares: snapshot a scenario on one set, press the other button (or move the equity returns a point), and compare the two success percentages — the difference between two plans is more trustworthy than the level of either (Monte Carlo and scenarios). A plan that swings 15 points on one point of return is telling you something a better forecast wouldn't.
• Enter what you know; shade what you don't. Fees you can look up, so type them rather than shading returns for them. What you can't look up is how the first decade goes, and the engine has one return per asset class forever — which is why a new plan starts on Conservative, as if that decade lasts. Set to Consensus, compare, and read the gap as the size of that risk.
• The remaining edges have no field at all. No correlation matrix, no fat tails, no glide path — each is on the list of what's left out. The first two are already charged inside the volatility the engine simulates (the straight-line blend, then the 1.05); the rest is what the two sets and the target below are for.
• The other cushion is the target. Aiming for a higher projected success than you strictly need buys room for these assumptions to be wrong — see Projected success.

Sources

• Elroy Dimson, Paul Marsh and Mike Staunton, UBS Global Investment Returns Yearbook 2026, public summary edition. US equities, bonds, bills and inflation, 1900–2025, annualized, nominal and real; developed-market equities in US dollars.
• J.P. Morgan Asset Management, 2026 Long-Term Capital Market Assumptions (October 2025). US-dollar assumption matrix: compound return, arithmetic return and annualized volatility per asset class; 10–15 year horizon.
• Horizon Actuarial Services, Survey of Capital Market Assumptions, 2026 Edition. Exhibit 17: the equal-weighted average of 43 advisors' assumptions; 20-year figures from the 29 who gave them; arithmetic, geometric and standard deviation. A new edition appears each summer, and Consensus is re-rounded from it.
• Vanguard, Vanguard Capital Markets Model forecasts, June 30, 2026 run. 10-year annualized nominal return ranges. Vanguard republishes these through the year.
• Michael Kitces, "Volatility Drag: How Variance Drains Investment Returns", Nerd's Eye View, December 2017.
• Federal Open Market Committee, Statement on Longer-Run Goals and Monetary Policy Strategy, as reaffirmed January 27, 2026.
• Federal Reserve Bank of St. Louis, 10-Year Breakeven Inflation Rate (T10YIE).
• Peterson-KFF Health System Tracker, "How does medical inflation compare to inflation in the rest of the economy?", August 2024.
• Fidelity Digital Assets, "A Closer Look at Bitcoin's Volatility", May 2024.

The outside figures on this page are dated facts about other people's studies; they will age, and the dates are there so you can tell. The defaults, and every figure derived from them, are checked against the engine each time the app is built.