The nuts and bolts of LILI We used the OECD database of monthly economic indicators to look for non-financial variables that could anticipate the CLI. We use small and simple forecasting regressions, as they sometimes adjust more quickly to structural changes than large regressions or regressions based on large data sets. And, we already have a big-data leading indicator, published in our Year Ahead a year ago. We looked only at models with two independent variables (and their lags) and with variables of the same “type” (production variables, or employment variables, or confidence indicators). Further research can be done to use models that mix variables, although those models are more difficult to interpret. We used two criteria to select between models. One was Granger-causality tests, which amount to joint tests that the lags of the independent variables are statistically different from zero in our dynamic regressions. The other was the Bayesian Information Criteria (BIC), a measure that looks at the fit of the regression (the r-squared) but that penalizes large models. We found that production measures and confidence indicators usually performed better in our training sample (1980 to 2014), although confidence measures usually had a larger lead. The model that used has a lead of four months and uses up to four lags of the standardized consumer confidence indicator (CCI) and of the standardized business confidence indicator (BCI) calculated by the OECD. Many models showed similar performance, which indicates that a combination strategy could be fruitful, although we did not take that route. No out-of-sample tests were performed. We like LILI because it can anticipate the OECD’s leading indicator without the use of financial variables. But we also like it because it is easy to interpret as it is based on confidence indicators and because those indicators are not subject to data revision, unlike variables such as GDP or payrolls. 12 Global Rates, FX & EM 2017 Year