We Tested Google's Forecasting AI on 5 Housing Markets: Right in Greece, Wrong in Portugal

We hid the latest 2 years of house prices in 5 markets and let Google's TimesFM 2.5 forecast them blind. It was 0.9% off on average in Greece, but in Portugal prices rose 37% while the model saw 3.3%. The AI keeps pace with steady markets and falls behind booms: one input, never a promise of returns.

Issue 1: Can AI forecast house prices?

Even Google Sheets can now call TimesFM, but for a family buying abroad, the useful question is how often the AI has been right before.

TimesFM 2.5, Google Research's time-series model, has public code on GitHub and Apache-2.0 weights that allow commercial use. We took the quarterly nominal house price index from the Bank for International Settlements (BIS) for Greece, Malaysia, Australia, Spain and Portugal, hid the latest 8 quarters and let the model forecast them untrained. It also had to beat 2 pencil-and-paper rules: hold the last price flat, or extend the latest 4-quarter trend.

The answer is yes, but only in 2 of 5 markets. Greek prices rose 14.0% over the test and the model forecast 13.4%: an average error of 0.9% against 2.7% for the trend rule, with 7 of 8 real quarters inside its 80% range, the band between its low and high cases. Malaysia's error was 1.4%, level with the trend rule, though the model ran hot at 7.3% against a real 4.0%.

The other 3 missed badly. Australian prices rose 12.7% while the model forecast a 0.7% fall. From Q2 2024 to Q1 2026, Spain rose 26.7% and Portugal 37.0%; the model saw 4.3% and 3.3%, and in Portugal not one quarter fell inside its range.

Issue 2: Why the model misses booms

A statistical model extends past patterns, so it tends to dampen sharp accelerations: the longer a market runs hot, the longer the model keeps forecasting a slowdown that hasn't come. Greece, up 68.6% from Q3 2020 to Q2 2026 with barely a stumble, is the steady kind of series it handles best.

The telling detail sits in the 3 misses. In Portugal, Spain and Australia alike, simply extending the 4-quarter trend beat the AI: 9.2% error against 12.9%, 6.2% against 9.6%, and 3.4% against 5.1%. The simple rule still fell well short of real prices, so nothing in the test saw the boom coming.

In all 3 boom markets, a simple trend rule still beat the AI.

We had already seen the same weakness on climate. On the monthly global temperature series from NASA GISTEMP, TimesFM forecast an average anomaly of 0.93 °C for 2021–2025 against an actual 1.14 °C, and lost to a plain 30-year trend line.

Issue 3: How to read the 2027–2028 scenarios

Only Greece and Malaysia passed the test, so only they get scenarios to Q2 2028. These are model scenarios – not NAC's forecast, and certainly not a promise about prices.

For Greece, the middle path takes the index from 122.4 to 128.8, a 5.3% rise over 2 years and a clear slowdown from the 5.5% gain of the latest 12 months alone. Yet even in Greece, where the model did best, its 80% range runs from a 7% fall to a 17.3% rise.

Even in Greece, where the model did best, its 80% range runs from a 7% fall to a 17.3% rise.

A band that wide is the honest answer: prices could go several ways, and one tidy number in the middle would hide that.

Malaysia's middle path rises 2.3%, within −1.8% to +6.2%. The band is narrow because prices there have moved slowly and evenly, but the model overshot in the test and caught only 5 of 8 quarters, so this band may be too narrow.

The BIS index is also a national average in local currency, not a single project's price, and one 8-quarter test isn't enough to say the model will be right next time.

Issue 4: How families should use AI forecasts

Treat a forecast as one input, not the answer. Weigh the price scenario against rental yield, costs, tax, the exchange rate and your family's residency goal – a Golden Visa home such as Keranis Ports Residences in Piraeus, say, or an MM2H base like Pavilion Square in Kuala Lumpur.

Read the whole range. Greece's +5.3% sits inside a band from −7.0% to +17.3%, so work out what your family could absorb if prices landed at the low end.

Be most careful in hot markets. There, price per square metre, rental yield and supply outweigh a forecast line; the NAC Quarterly Report for Q3 2026 tracks the wider picture.

Frequently asked questions

Does Greece's +5.3% scenario mean prices will rise 5.3%?

No. It's the model's middle path for the national average index, inside an 80% range from −7.0% to +17.3%, and prices in a given district or project can differ widely.

Why are there no scenarios for Portugal, Spain and Australia?

Because the model failed the test in those 3 markets, and publishing scenarios from a model that just missed badly wouldn't be honest.

NAC strategic insight

AI is good at extending the past and weak at turning points – exactly when investors want a forecast most. We put price scenarios next to yield, costs, exchange rates and the residency route, so data sharpens the questions rather than promising a number. To see how that fits your family's plans, book a consultation with NAC.

Sources

• BIS — Residential property prices, quarterly nominal index (2026)
• Google Research — TimesFM on GitHub: versions 2.5 and 3.0, weight licences (2026)
• Google Research — A decoder-only foundation model for time-series forecasting (2024)
• Google Research — TimesFM-3: A zero-shot foundation model for multivariate forecasting (2026)
• NASA GISS — GISTEMP global surface temperature analysis (2026)