Forecasting the Dragon: Calibrating Superforecasting Against Bloomberg's China and Midterm Data
Tetlock and Gardner's central claim is not that some people can see the future. It is that most of us are worse forecasters than we think, that a small number of ordinary people can be trained to be reliably better, and that the difference is measurable — in Brier scores, in calibration curves, in years of tournament data. This post takes that claim out of the IARPA tournament and into two forecasting arenas Bloomberg covers every week of 2026: China's GDP prints and the U.S. midterm generic ballot. The question is whether the fox-versus-hedgehog distinction that explains superforecaster performance in a controlled tournament also explains why professional consensus forecasts on China have been missing by widening margins all year, and why Bloomberg Politics keeps hedging its midterm coverage in probabilistic rather than declarative language.
I.The Fox, the Hedgehog, and the Ten-Point Scoring Rule
The intellectual spine of Superforecasting is borrowed from Isaiah Berlin's essay on Tolstoy and repurposed as an empirical hypothesis. Hedgehogs organize the world around one big, elegant idea and force new information to fit it; foxes hold many small models loosely, update them incrementally, and are comfortable saying "I'm 65 percent confident" instead of "this will happen." Tetlock's original research in the 1980s and 1990s — long before the book — found that expert political forecasters performed barely better than chance, and that the most famous, most frequently quoted experts were often the worst performers, because television and op-ed culture rewards confident, hedgehog-style narrative over careful, foxy hedging.
The book's real contribution is what came after that discouraging finding: the Good Judgment Project, an IARPA-funded tournament that recruited thousands of volunteer forecasters, scored their predictions against real geopolitical and economic questions using Brier scores, and discovered that a small subset — the superforecasters — consistently outperformed not only the volunteer pool but, by the tournament's own account, intelligence community analysts working with classified access, by a margin regularly cited at around 30 percent. What distinguished superforecasters was not raw intelligence but process: breaking big questions into smaller sub-questions, anchoring on base rates before adjusting for the specific case, updating in small increments rather than lurching, and treating every prediction as a probability rather than a verdict.
What the book does not spend much time on — and what this post uses as its pivot — is institutional forecasting: the kind Bloomberg Intelligence, Bloomberg Economics, and the wire desks that make up Bloomberg Politics produce every day under deadline, for an audience that wants a number, not a probability distribution. That is a much closer analogue to how forecasting actually gets consumed in markets and in politics than the GJP tournament was, and it is worth asking whether the same foxy disciplines hold up under commercial and editorial pressure.
II.Case One: China's GDP Consensus as an Unplanned Forecasting Tournament
Bloomberg runs a standing survey of economists on China's quarterly GDP, and 2026 has produced an unusually clean natural experiment in calibration. Beijing itself behaved like a chastened forecaster this year: in March, the government lowered its 2026 growth target to a range of 4.5 to 5 percent — the first formal downgrade since 2023, and the least ambitious official target since 1991. A range instead of a point estimate is, in Tetlock's terms, an admission that the underlying system is harder to read than the old single-number targets implied. That is itself a foxy move by a historically hedgehog-style planning apparatus.
The quarterly data tested that admission almost immediately. Q1 2026 GDP grew 5.0 percent year-on-year, beating the Bloomberg-surveyed consensus of 4.8 percent — a miss in the optimistic direction. Q2 2026 came in at 4.3 percent, below both the bottom of Beijing's own target band and the Bloomberg-polled consensus of 4.5 percent — a miss in the pessimistic direction, and a sharper one. Two quarters, two misses, in opposite directions, is exactly the pattern a well-calibrated but genuinely uncertain forecasting process should produce; a badly calibrated hedgehog process, by contrast, tends to miss in the same direction repeatedly because it is anchored to a single fixed narrative (secular Chinese deceleration, or resilient export-led growth) rather than updating quarter to quarter.
The forecasting houses split along a genuinely foxy-versus-hedgehog line here. Goldman Sachs published an explicitly out-of-consensus call — above-consensus growth of 4.8 percent for the year, built on a specific, falsifiable mechanism (a rising current-account surplus driven by export resilience to emerging markets) rather than a general China narrative. UBS and BBVA Research, by contrast, held closer to the Bloomberg consensus median near 4.5 percent, citing the more familiar structural story of a fifth straight year of property-sector contraction dragging against export strength. Neither camp was "wrong" in the hedgehog sense of having an unfalsifiable thesis — both published specific numbers that were later scored against reality — but the dispersion itself is the data point Tetlock would want us to notice: a 2026 China consensus that is officially a range rather than a point, and a set of bank forecasts still clustering within half a percentage point of each other, is a market that has partially absorbed the calibration lesson without fully abandoning point-estimate culture.
The methodological upshot: a forecast that misses in alternating directions across consecutive quarters is not proof of skill, but it is evidence against the specific failure mode Tetlock spent two decades documenting — a fixed narrative that gets defended against disconfirming data rather than updated by it.III.
Case Two: Bloomberg Politics and the 2026 Midterm Calibration Problem
The second tournament is louder and more partisan, which is exactly the environment Tetlock's original research identified as hardest for good forecasting to survive in. Bloomberg's own midterm coverage through 2026 has tracked a consistent structural setup: a president with approval in the high 30s to low 40s, a generic congressional ballot favoring Democrats by roughly five to eleven points depending on the poll and the likely-voter model applied, and a House majority thin enough that Democrats need a net gain of only three to five seats to flip control. Historical base-rate models — the kind Tetlock explicitly recommends starting from before adjusting for case-specific detail — point toward the president's party losing House seats in almost every midterm below a 46 to 47 percent approval threshold, which would suggest a Democratic-favorable outcome on base rates alone, before anyone reads a single fresh poll.
What is analytically interesting is less the direction of the forecast than how it is being communicated. Bloomberg's coverage through the year has consistently hedged: describing Democrats as holding an "early advantage," flagging that momentum has to be sustained for ten months to become a wave, and treating individual outlier polls (an 11-point Democratic lead in one late-cycle survey against a 7-point lead in the same pollster's registered-voter sample) as data to be reconciled rather than cherry-picked. That is foxy behavior in Tetlock's sense — updating incrementally on noisy data rather than locking onto the most dramatic single poll — even though the underlying medium (daily political journalism) is exactly the genre his original research found rewards hedgehog overconfidence.
What a Calibration Table Would Ask
Tetlock's preferred evaluation tool is not "did the favorite win" but a calibration table: among every race a forecaster called a 55 percent toss-up, did the favored side actually win about 55 percent of the time — no more, no less? Election forecasters who publish this kind of scorecard after past cycles have generally shown a mild, structurally interesting pattern: high-confidence calls (races rated above 95 percent) are called correctly essentially all the time, while true toss-ups — the 50-to-60-percent band — perform close to a coin flip, exactly as calibration would predict, but with real variance year to year because the sample of genuine toss-ups in any single cycle is small. The lesson for 2026 is not to trust any single-point prediction about which chamber flips, but to ask whether the institutions making that call are willing to publish their toss-up band honestly rather than rounding every race into a headline-friendly certainty.
A Simple Cross-Domain Calibration Table
Laying the two cases side by side gives a small but genuine calibration exercise — not a rigorous Brier score, which requires many repeated trials, but a first-pass diagnostic of the kind Tetlock recommends before trusting any single forecasting house's track record.
| Forecast | Consensus / Source | Outcome | Miss | Direction |
|---|---|---|---|---|
| China GDP, Q1 2026 | 4.8% (Bloomberg survey) | 5.0% (NBS) | +0.2 pt | Optimistic miss |
| China GDP, Q2 2026 | 4.5% (Bloomberg survey) | 4.3% (NBS) | −0.2 pt | Pessimistic miss |
| China GDP, FY2026 | 4.5–5.0% (official band) | 4.7% (H1 actual, pending) | Within band | Band held |
| 2022 midterm toss-ups (50–60%) | Favorite-side odds | ~54% actual win rate | Near-calibrated | Reference case |
Two things stand out. First, the China consensus missed in both directions across successive quarters by roughly the same magnitude — the signature of noisy-but-unbiased forecasting rather than a systematically hedgehog-biased one. Second, the one calibration table available from a comparable past electoral cycle shows toss-up races landing almost exactly where their stated odds implied, which is the single strongest piece of evidence in Tetlock's favor: probabilistic forecasting, honestly reported, really does behave the way the math says it should, even in the adversarial, high-emotion environment of electoral politics.
V.Where the Framework Strains
An honest academic treatment has to flag where Superforecasting's claims travel less well outside the tournament that generated them, and this blog's stated commitment — treating a framework as an argument to be tested rather than a lens to be applied uncritically — applies here as much as it did to the Trojan Economy series.
- Short horizons versus strategic horizons. The GJP tournament scored questions with resolution windows of months, not years. Tetlock himself has been candid that superforecasters' edge shrinks on multi-year, structural questions — precisely the kind this blog's prior posts on China's engineering state and the Belt and Road chokepoint network are built on. A quarter-ahead GDP call and a decade-ahead judgment about whether an "engineering state" model outcompetes a "lawyerly society" are not the same forecasting problem, and conflating them risks importing false confidence from the tournament's track record into domains it was never tested on.
- Adversarial opacity. Chinese macroeconomic data and Communist Party political intentions are, by design, harder to observe than an American election's public polling. Superforecasters in the IARPA tournament operated on open-source information about mostly transparent systems; applying the same confidence to a system with an actively managed data apparatus understates a genuine information asymmetry that no amount of base-rate discipline fully corrects for.
- Incentive contamination. Bank research desks and campaign-adjacent forecasters are not disinterested tournament participants. Goldman's above-consensus China call and any bank's midterm-adjacent political risk note both carry commercial incentives — to be memorably right, or to hedge exposure — that the anonymous volunteer forecasters in Tetlock's tournament did not face. Calibration measured against a track record with unclear incentive structure is weaker evidence than calibration measured in a blind, incentive-neutral tournament.
- Survivorship in the retrospective table. The one calibration table cited above comes from a single forecaster's single past cycle — encouraging, but not itself a large-sample validation, and it is presented here as illustrative rather than as proof that 2026 forecasts will calibrate the same way.
Synthesis: What This Adds to the Series
Read together with this blog's earlier work, the picture that emerges is of two different kinds of institutional forecasting culture sitting on either side of the U.S.-China relationship, and Tetlock's fox/hedgehog distinction turns out to be a genuinely useful diagnostic for both — not because either side has become a tournament-grade superforecaster, but because both have visibly moved toward foxier habits under pressure. Beijing's shift from a point target to a range is a small, forced concession to uncertainty from a planning tradition built on the opposite instinct. Bloomberg's midterm coverage, operating inside a media genre Tetlock's original research found actively punishes hedging, has nonetheless leaned on probabilistic, base-rate-anchored language rather than declarative certainty. Neither is proof that either institution has internalized the ten commandments for aspiring superforecasters wholesale. But the pattern is consistent with the book's more modest and more defensible claim: calibration is a trainable habit, visible in how confidently an institution is willing to state a number, and the 2026 data — on both the Chinese economy and the American electorate — rewards the institutions doing the hedging over the ones still selling certainty.
This post is offered as argument, not as a verdict on either China's 2026 growth path or the midterm outcome — in keeping with the probabilistic spirit it is arguing for. Readers are invited to treat every number above as a forecast still open to updating, not a conclusion.
Comments