EXPERIMENTALExperimental model. These models rest on a lot of assumptions: future earnings growth, where valuation multiples end up, the return each risk should pay, volatility taken from past years, and a constant EUR/USD rate. Change any of them and the ranking can change. Read the numbers as scenarios, not forecasts. Not investment advice.
ETFs vs AI stocks: one Monte Carlo model
We put €10,000 into each of twenty-six choices and ran 20,000 simulated futures for each: eight index ETFs (world, US, Europe, emerging markets, the US tech sector, an AI fund and a chip fund), ten AI stocks, bitcoin, gold, a rented new-build flat in Bratislava bought with cash or with a mortgage, the 30-year US inflation-linked Treasury (TIPS), and, as the yardstick, German government bonds (Bunds). Over 5 years the broad funds give about what Alphabet and Nvidia do, with much smaller falls on the way. Only SK hynix and bitcoin clear the ~9.7% a year that stock risk should pay today, but with both, in the middle future you sit through a fall of almost 60% on the way.
Holding period Click a column to sort
€10,000 held for 5 years: return, risk and Sharpe
Shaded rows are index ETFs. Hover or tap the orange ! by a flat for every input behind it. German Bunds, the yardstick, stay at the bottom and are left out of the best/worst marks. Green is the best value in a column, orange the worst. Returns in EUR at constant exchange rates.
A share price is roughly earnings × P/E. So over 5 years the return is about earnings growth + the change in P/E, plus dividends and buybacks for the funds. Middle (median) future of each model.
“P/E as quoted” is price ÷ the next 12 months’ expected earnings, the number you see on finance sites. The model’s own P/E adjusts it: net cash is taken out of the price (net debt added), and SK hynix and the chip ETF’s memory makers are measured on normal-year earnings, not today’s peak. SK hynix’s model P/E drifts toward 11–17× those earnings, the top of its own record; the small line under it reruns the same futures at the ~9× (7–11×) that past memory cycles paid. ASML and TSMC have the same line: their model P/E drifts toward an AI-era 20–28× and 13–19×, and the small line reruns the futures at what the price paid on earnings that came true in 2010–2019, about 17× (13–21×) and 11× (9–13×). Index funds are modelled by region on today’s CAPE or trailing P/E, which drifts back toward its long-run norm.
Margin-of-safety price: how low would it have to be?
The highest price at which buying passes our safety rule. In at least 8 of 10 futures, both must hold: (1) at the end you beat a government bond of the same length, in the same currency; (2) on the way you never fall more than the fall limit below what you paid. The fall limit is 1.5 × the yearly volatility, kept between 20% and 50%.
Expected return vs the typical fall on the way, 5 years
Up is more return; left is a deeper median drawdown. Top-right is where you want to be.
EV p.a.: the average outcome of all 20,000 futures, as a yearly rate. Big wins in a few futures pull it up.
Median p.a.: the middle future. Half do better, half worse. For a single stock it is often well below the EV.
P(loss): the share of futures that end below €10,000.
Worst 5% / best 5%: what the €10,000 is worth at the end, in a bad (1 in 20) and a good (1 in 20) future.
Median max drawdown: the deepest fall from a high you would sit through along the way, in the middle future.
P(fall > 50%): the chance that at some point the position is worth less than half its previous high.
Sharpe: return above EUR cash (2% a year) per unit of risk. It is the fairest single number to compare a stock with a fund. A Bund held to maturity is the exception: its result is known in advance, so its 5-year Sharpe is huge. It is the yardstick, not the winner.
What we read from it
Method
One set of futures. GOOG, NVDA and AMD are modelled business by business (search, cloud, data-centre chips and so on) on shared drivers: AI spending, the economy, interest rates. META and MSFT use a lighter version on the same drivers: two years of analyst earnings forecasts, then a fade. Dell is modelled line by line: AI servers (the $74bn guided for this year and the $95bn order backlog, then the AI market with a slow loss of share to Taiwanese contract makers, and a thin margin), traditional servers and storage (inflated by memory prices today, partly unwinding), and PCs. Its multiple drifts toward that of a hardware maker (11-17x, as Dell, HP and HPE traded before the boom). INTC comes from our semiconductor model (its foundry is an option, not the base case). ASML is modelled as lithography systems (supply-bound to 2028, then growing with chip makers' spending, cut 25% in a digestion year, with a chance of tighter China export rules) plus its service business, at an operating margin of 42-50%. TSMC is modelled as AI accelerator wafers (GPUs and custom chips alike) plus everything else, at a net margin that fades from 51% to 40-50% as overseas fabs and the capex wave are paid for. Both start from analyst forecasts to 2028. TSMC also carries a 1.5% a year chance of a Taiwan Strait blockade or worse, which leaves about a third of its value; ASML, whose biggest customer it is, keeps about three quarters in the same futures. The other stocks are not hit in those futures, though Nvidia, AMD and Apple would be too. SK hynix is modelled as a memory cycle: the 2026-2027 boom, then normal-year profits of 20-50% of the 2027 peak, plus its net cash. Shared drivers mean the stocks fall together when AI spending slows, as they would in real life.
Bitcoin comes from our separate bitcoin model, on its own 20,000 futures, so it is not tied to the AI drivers. It blends four trend models by how well each forecast 1-5 years ahead in a walk-forward test since 2013: a power law fitted on all history (25%), a maturity model in which the power law can bend and volatility shrinks with age (50%), decaying growth (12.5%) and a hand-written scenario table (12.5%). The ups and downs are resampled from the last 5 years of weekly returns (about 50% a year), and each year carries a 1% chance of a permanent −80% break.
Gold moves as it has in EUR since 1971 (resampled in 24-month stretches, about 17% volatility a year), held as a physical-gold ETC (0.12% a year fee). Gold has no earnings to anchor it, so its expected return is a scenario table we chose and show: a 30% chance that the dear price reverts as the 1975–2021 pattern says (about −7% a year after inflation), 40% that gold earns its long-run +1.6% a year after inflation, and 30% that central banks keep buying as they have since 2022 (+5% after inflation). That averages to about 0% after inflation, and each future draws its own return from that spread (about 4.8% a year). Each scenario alone is shown under its row in the P/E table. The reversion pattern held from 1975 to 2021, but so did gold's link to real interest rates, and that broke in 2022. Physical bars are tax-free in Slovakia when sold at any time; whether an ETC passes the one-year test for securities is not settled.
The flat is a 2-room new build in Ružinov or Petržalka (not the centre), priced from listings for sale and for rent on Reality.sk on (2-room, 38–65 m², new build, duplicates removed), at the median asking price per m² and the median rent without energy. They match Realitná únia's district averages and developers' price lists. Prices move as they have in the Bratislava region (NBS, 2005–2021) and Slovakia (Eurostat, 2022–2026), resampled a year at a time and corrected for the smoothing in such indices: about 12% a year instead of the measured 6%. The price trend is a stated view, +3% a year; the flat also ages against the index and sells 5% above or below it by chance. The owner pays the repair fund, insurance, property tax and upkeep; the tenant pays energy. Rent is taxed at 19% after costs and interest; a sale after 5 years is tax-free, a sale before that pays 35% of the gain. With a mortgage, the monthly shortfall is borrowed at 3%, so the result is per €10,000 of the money put in on day one. Hover the ! in the table for every number.
German Bunds are bought at par today with a coupon equal to the yield, from the Bundesbank's curve for federal securities ( average: 5-year , 10-year ). Each month the bond is repriced on that curve plus moves resampled from the euro era (1999 on), with their average removed, so rates are as likely to rise as to fall. The 5-year is held to maturity; the 10-year is sold after 5 years at the then 5-year yield, so its result depends on where rates go. Coupons are taxed at 13% in Slovakia (the separate tax base for government bond income, from 2026) and held in cash; a sale after a year is tax-free. An older Bund with a low coupon bought below par turns most of the return into a tax-free price gain, so it pays a little more after tax.
The US TIPS is the 30-year inflation-linked Treasury of February 2056, bought at above US inflation (), from our TIPS model. Each month its real yield, US inflation and the EUR/USD rate move together as they did since 2000 (resampled in about 24-month stretches), with the real yield held level on average and the dollar losing the gap between US and euro inflation. It is held unhedged, so unlike the stocks it carries the dollar's swings. Coupons are taxed at 13%; the sale after 5 years is at the then real yield.
Every stock model includes a 15% a year chance that AI spending pauses (a "digestion" year). This is a judgement, and it moves every stock number. The ranking between the names holds when we change it; the levels do not.
Index ETFs (All-World, MSCI World, S&P 500, Europe, emerging markets, S&P 500 IT, XAIX, and the chip ETF in its index version) start from today's valuations: the expected real return is dividends plus net buybacks plus real earnings growth, plus the drift as today's P/E moves toward its long-run norm, plus 2.1% inflation, ±1.5% per future. The ups and downs are resampled in 24-month blocks from each fund's own EUR history: world since 1972, S&P 500 since 1988, chips since 1994, S&P 500 IT since 1999 (XLK, so the 2000-02 tech crash is in), emerging markets since 2003 (EEM), Europe since 2005 (VGK), and XAIX since 2019, extended back to 2001 with global tech (IXN), which it tracked best.
The sector funds rest on judgement. S&P 500 IT and XAIX assume tech earnings keep growing 3.5-4% a year above inflation, and that their P/E (32 and 26 today) drifts toward 24 and 21 with a 10-year half-life. Change those and their rows move more than any other index row.
The chip ETF twice. SEMI (iShares MSCI Global Semiconductors UCITS) is shown under the index model and under a bottom-up model of its holdings. The bottom-up version values memory makers on normal-year profits, not today's peak, which is why the two disagree.
Prices are closes on (SK hynix: KRW , Seoul close on ; Dell: $ on ; ASML: € in Amsterdam and TSMC: NT$ in Taipei, on , where the TSM ADR in New York costs about 10% more for the same shares; bitcoin: $ on ). Returns are in EUR at constant exchange rates. In Slovakia, gains on shares and ETFs held over a year are tax-free, so the figures are also after tax. Bitcoin is shown the same way, as an exchange-traded product held over a year; whether such a product passes that one-year test is not settled. Coins held directly are taxed on every gain since 2026, with no time test, which takes about two fifths off the 5-year expected gain.
Required return ~9.7% a year: the 10-year US Treasury (5.2%) plus a 4.5% equity risk premium.
Not investment advice. These are model outcomes with judgement inside them. They tell you how the choices compare on equal terms, not what will happen.