GDP nowcast update

Summary

I've been working on a GDP nowcast over the last few months. Three models now, all pointing at a hot 2026 Q1. The rest of this post is about why I don't fully believe them.

I foreshadowed this work back in April with a Bridge model write-up. The Bridge has been refined since then. It now runs 13 bridges rather than the 7 in the April post (private capex and construction split out as their own bridges, the NAB conditions survey added, household spending brought in alongside the existing consumption bridge), and most of the within-quarter data is in. The remaining pieces, business profits and government final consumption, are out next Tuesday, one day before the national accounts. So this isn't quite the final pre-release nowcast, but it is close.

In addition to the Bridge model I now have a dynamic factor model (DFM) and a Bayesian Vector Auto-regression (BVAR) model. Three independent ways of producing a nowcast running on a similar input panel.

Here is where they land for 2026 Q1.

Model	QoQ %	TTY %	70% CI (QoQ)	90% CI (QoQ)
Bridge	+0.82	+2.96	[+0.54, +1.12]	[+0.38, +1.30]
DFM	+0.86	+3.01	[+0.20, +1.53]	[−0.19, +1.92]
BVAR	+0.80	+2.95	[+0.22, +1.39]	[−0.12, +1.73]

The three models agree on a print near +0.8 QoQ, roughly +3.0 TTY. The Bridge runs the narrowest band because it conditions on monthly indicators directly. The DFM and BVAR widen as expected given their factor and VAR formulations.

What the bridges are saying

Before getting to the diagnostic, it is worth pausing on the dispersion across the individual bridges, because the headline +0.82 is a weighted average of bridges that genuinely disagree about the quarter.

Labour hours is at +1.24 percent QoQ. The Cobb-Douglas production tracker is at +0.05 percent. That is a 1.2 percentage point spread inside one model. The demand-side bridges (labour, investment, prices, trade) are mostly clustered between +0.6 and +1.0. The supply-side anchor is barely moving. The combined estimate is averaging across that disagreement rather than resolving it.

The weights matter for what comes next.

Labour bridges (hours and employment) carry 42 percent of the combined weight. The investment block (private capex plus construction) carries about 11 percent. So the headline is being pulled to +0.82 mainly by labour, with investment corroborating rather than driving. That matters for the diagnostic argument below, which is about whether the investment signal can be trusted.

A three-body problem

The GDP identity sits underneath all three models. GDP is Consumption + Investment + Government + eXports − iMports, and during an import-funded investment cycle, that identity gets awkward for nowcasts.

$$ GDP = C + I + G + X - M $$

A worked example. If a data centre operator in Sydney buys \$100 million of GPUs from the US and installs them, equipment capex (part of I) rises by \$100 million and goods imports (M) also rise by \$100 million. In the GDP sum, the +100 in I and the −100 in M cancel, and the contribution to Australian GDP is zero. Which is correct, because Australia did not produce anything. It just moved a pile of foreign-made silicon from a ship to a server rack.

If the same operator had bought \$100 million of GPUs from a factory in Geelong, the I line would still rise by \$100 million but M would not move. Net contribution to GDP would be +100. Also correct, because Australia did produce those GPUs.

The identity makes sure you cannot fake domestic production by importing things. The trouble for a nowcast is that the three lines that have to balance, capex, imports, and inventories, are not equally visible in real time. Capex is clean. Goods imports is visible with a lag and a goods-only restriction. Inventories, as a national accounts concept, is essentially invisible until the GDP release itself.

The capex print

Equipment capex on a chain volume basis ran at +0.61 percent of GDP in QoQ terms this quarter, against a 1997-onwards historical mean of +0.03 percent and a standard deviation of 0.13. That is a 4.4-sigma move. The series has fat tails, so the sigma overstates the rarity a little, but it is the largest equipment-capex-as-share-of-GDP reading in 28 years of data, including the mining boom.

The AI buildout is the obvious candidate explanation. The major hyperscaler announcements for Australian data centres went through 2024 and 2025, and the equipment is being installed now.

The imports that should match it

Under the identity, an import-funded capex surge should see goods M rise alongside I. It is not really happening. Goods imports QoQ as a share of GDP came in at +0.44pp, against a historical mean of +0.17 and a standard deviation of 0.42. That is only +0.6 sigma. The capex is 4.4 sigma hot, the imports are barely moving.

There are at least four reasons the offset might not be where you would expect to find it.

• Timing. The capex survey records when businesses booked the spend, while the trade data records when goods cleared customs. For lumpy items like data centre equipment, those can be a quarter or more apart in either direction.

• Services imports. A meaningful chunk of AI buildout spending is cloud contracts, software licences, and capitalised digital services, which flow through services debits rather than goods debits. My diagnostic only looks at goods.

• Domestic content. Some of the capex really is domestic. The shell, the fit-out, the electrical and cooling work. That portion legitimately adds to GDP with no matching import.

• Inventories. This is probably the biggest one. If the equipment cleared customs in earlier quarters and went into inventories, the goods M hit already happened, and what we are seeing now is an inventories drawdown into installed capex that nets out within the I block.

The corner nobody can see

The inventories possibility is what nags at me, because it points to a structural problem rather than a calibration issue. Capex I can see. Goods imports I can see (with a lag, and goods only). Inventories as a national accounts concept does not exist outside the 5206.0 release. The ABS compiles it during the national accounts process and you see it when you see the GDP print.

So the three corners of the triangle that the GDP identity binds together, capex and imports and inventories, are not equally visible to anyone trying to nowcast the quarter. Two are visible in real time with caveats. One is not visible at all until the moment GDP itself is published.

More data lands on Tuesday. Business profits, sales, inventories, and government final consumption all print one day before the national accounts. The Bridge picks up all three of the business indicators series. The DFM picks up only profits. The BVAR doesn't move at all on Tuesday, since none of these series are in its panel. But none of this resolves the structural problem. Even with the Tuesday data, inventories as a national accounts concept doesn't exist until GDP itself is published. So I started building a diagnostic that sits alongside the three models, doesn't try to nowcast anything, and just measures the gap between what the models can see and what the identity will eventually impose.

For 2026 Q1 the diagnostic puts the gap at +0.31pp in capex-deviation minus imports-deviation as a share of GDP. That is not the same thing as +0.31pp of nowcast bias, since the investment block in the bridge only carries about 11 percent of the weight and the labour bridges (which would survive an I/M netting) are doing more of the work. So the actual upside bias in the +0.82 print is probably closer to one or two tenths of a percentage point. That depends on how much of the labour signal is itself reflecting buildout activity that will get netted out elsewhere in the accounts. The diagnostic is flagging that something is off. It is not pricing an exact correction.

The DFM says nothing unusual

The DFM is doing something worth flagging too.

Both factors spiked through COVID and have settled back to roughly normal range from 2024 onwards. The current values are essentially noise. So the DFM isn't reading the economy as being in an unusual factor configuration. The +0.86 it's printing is coming through ordinary loadings on a fairly ordinary-looking factor state, not from the factors themselves saying the world has changed. Combined with the Bridge and BVAR landing in the same place, the three-way agreement looks more like the three models reading the same indicator surface than three independent confirmations.

What to look for in the release

When 2026 Q1 GDP lands, three scenarios.

If the print comes in near +0.8, the capex was genuinely domestic and the bridges were right. That would itself be interesting, because it would say the AI buildout is more locally additive than the import-heavy narrative suggests.

If the print comes in closer to +0.6 with inventories drawing down sharply, the offset was hiding in inventories. The bridges over-read the capex because they could not see the stock that was being depleted to fund it.

If it comes in closer to +0.6 with M revised higher or with services imports surging, the offset was in trade flows the diagnostic could not fully see. Same conclusion, different mechanism.

I think the second is most likely, but I would not bet heavily.

Where this leaves me

The headline is +0.8 across three models. The diagnostic suggests the actual Q/Q GDP growth is a touch lower once the identity closes. Potential itself is now around +0.5 QoQ, so even a corrected print is above potential, just less spectacularly than the raw nowcasts suggest.

Tuesday's release will mostly test the Bridge's read. The inventories bridge in particular will do work the other two models can't. The full national accounts come a day later.

Whichever way it goes, the diagnostic is staying in the pipeline. Any quarter where one corner of the capex-imports-inventories triangle moves hard, the bridges will over-read in the direction of whichever corner they can see most cleanly. The AI buildout is the immediate occasion. The geometry is the permanent feature.