olivlaw

From books

The macro consensus does not see recessions: they appear in data late, forecasts are revised slowly, they are dated months later, and relations estimated on the past break when policy changes.

Prakash Loungani · How accurate are private sector forecasts? Cross-country evidence from consensus forecasts of output growth · 2001 · Prakash Loungani, «How accurate are private sector forecasts? Cross-country evidence from consensus forecasts of output growth», International Journal of Forecasting 17(3), 2001, pp. 419–432, rezumat (traducerea noastră)3 minutes read
“The record of failure to predict recessions is virtually unblemished.”Prakash Loungani · How accurate are private sector forecasts? Cross-country evidence from consensus forecasts of output growth · 2001 · Prakash Loungani, «How accurate are private sector forecasts? Cross-country evidence from consensus forecasts of output growth», International Journal of Forecasting 17(3), 2001, pp. 419–432, rezumat (traducerea noastră)

The consensus is a good anchor for figures published on a fixed date and a poor probability for turning points.

In 2001 Prakash Loungani measured (International Journal of Forecasting) consensus forecasts of GDP growth between 1989 and 1998 and concluded: ‘the record of failure to predict recessions is virtually unblemished’. Zidong An, João Tovar Jalles and Loungani took up the question again in 2018, for 63 countries and the years 1992–2014: forecasters know that recession years are different, but they get the size of the fall badly wrong until close to the end of the year, forecasts are revised slowly, and private and official forecasts are virtually identical. Two critiques explain why this is not mere negligence. Robert Lucas showed in 1976 that relations estimated on the past change when the policy rule changes, because people adapt; Charles Goodhart observed in 1975 that a statistical regularity used for control loses its link to what it measured. Data add to this: a recession is seen late. The NBER committee announced the December 2007 peak only on 1 December 2008, and the March 2001 peak on 26 November 2001, that is, after almost a year and after eight months respectively. The conclusion is not that macroeconomics cannot be forecast, but that turning points are the hardest to catch and the consensus is a weak starting point. A forecaster uses it as an anchor, not as a probability, and asks questions about figures published on a fixed date, not about turning points: there the Brier score quickly shows whether you added anything beyond the consensus.

Worked example: why catching a few recessions does not beat a flat forecast. The numbers are illustrative, built to show the arithmetic, not real data. In a sample of 200 country-years, 20 are recessions: the base rate is 10%. We compare two forecasters with the one-sided Brier score, (p − outcome)², averaged over cases.

The constant forecaster says p = 0.10 everywhere. For the 20 recessions, (0.10 − 1)² = 0.81; for the 180 years without recession, (0.10 − 0)² = 0.01.

Constant (p = 0.10):
  20 × 0.81 =  16.20
 180 × 0.01 =   1.80
  total        18.00   →  18.00 / 200 = 0.0900

The alarmist forecaster gives p = 0.8 in 8 country-years, of which 3 are recessions and 5 are not, and p = 0.05 in the remaining 192, of which 17 are recessions and 175 are not. It caught 3 of 20 recessions, that is 15%, with an alarm precision of 3 / 8 = 37.5%.

Alarmist:
   3 × (0.8 − 1)²    =  3 × 0.04   =  0.12
   5 × (0.8 − 0)²    =  5 × 0.64   =  3.20
  17 × (0.05 − 1)²   = 17 × 0.9025 = 15.3425
 175 × (0.05 − 0)²   = 175 × 0.0025 =  0.4375
  total                              19.10  →  19.10 / 200 = 0.0955

The forecaster who ‘sees’ recessions has a worse score (0.0955) than the one who sees nothing (0.0900). Against the flat version: the 3 recessions caught gain 3 × (0.81 − 0.04) = 2.31, but the 5 false alarms lose 5 × (0.64 − 0.01) = 3.15, and the 17 recessions missed at p = 0.05 lose 17 × (0.9025 − 0.81) = 1.5725; the 175 quiet years gain 175 × (0.01 − 0.0025) = 1.3125. Net: −2.31 + 3.15 + 1.5725 − 1.3125 = +1.10, that is 19.10 − 18.00. A signal that catches 15% of the turning points, with 37.5% precision, is not enough. That is why a consensus that misses turning points, as Loungani (2001) and An, Jalles and Loungani (2018) describe it on real data, adds little beyond the base rate precisely on recessions.

For the forecaster: a simple annual definition, used in the literature, calls a recession a year with negative real GDP growth; you can count by country and year from the IMF World Economic Outlook series or from national accounts (the National Institute of Statistics for Romania, Eurostat for the EU). The base rate that results is a better starting point than the consensus, especially for questions about turning points.

Indicators to follow: IMF forecasts in the World Economic Outlook (April and October) and those of the European Commission, with their history; the flash GDP estimate (the National Institute of Statistics about 45 days after the quarter, Eurostat about 30 days); the first release against the revised series for the same period; NBER dating, for the US.

The typical trap: scoring on revised data when the question should have been resolved on the first release, or the other way round. Revisions can move a small growth figure from one side of zero to the other. The second trap is treating the consensus as a probability; it is a central point, not a distribution, and in turning-point years it lies outside the range.

Example of a scorable question: ‘Will the flash estimate of Romania’s GDP growth in the second quarter of 2027, quarter on quarter, seasonally adjusted, be negative? Resolution: the National Institute of Statistics release published by 15 August 2027; later revisions do not count.’ The question is illustrative and gets no probability here; start from the base rate of negative-growth quarters in the INS series and then adjust with the data already published for the period.

Why it matters Loungani found that in 1989–1998 consensus forecasts missed practically all recessions, and the NBER announced the December 2007 peak only after almost a year.

Consensusstays nearSlowrevisions inFirst GDPestimate,NBER datesthe peak
At each step, information arrives later than the forecaster would need it.

← Back to the feed