Why do we nowcast?
Most EM governments publish data on debt cross-border flows with a one-to-three-month lag.
The objective of the nowcast models is to make monthly predictions of EM debt cross-border flows to countries with lagging data releases.
This way, we fill the missing debt flows data from recent months with estimates from our nowcast models.
We created nowcast models for each country, as well as for the total debt flows.
Each nowcast model is customised to obtain the best possible estimates.
All nowcast models leverage Trounceflow’s data on fund flows and debt flows from countries which publish their data with no lag, such as Indonesia, Hungary, Mexico, and Turkey.
The nowcast models are linear models with a dynamic approach.
For example, at the beginning of the month, debt flows data is only available for Indonesia and Hungary. When running the nowcast models at the beginning of the month, they will optimize based on Indonesia and Hungary debt flows, and other data available at that moment.
As more days pass, more inputs are added to the nowcast models, and their estimations’ accuracy increases.
How good are the nowcast models?
When nowcasting the flows for the current month, the prediction accuracy is around 40% on the first few days of the month, rising to 70-95% accuracy by the end of month.
The same accuracy of 70-95% is obtained for all the previous months for which data is yet to be released.
So, for example, if we were at the end of August, the nowcast models would estimate August, July and June with a 70-95% accuracy (for countries with a 3-month lag).
The usual progression of accuracy over a month is shown in the figure below.
What is the coverage (countries, timing, frequency)?
The nowcast models cover the countries in the GBI-EM index and the total EM debt flows.
The nowcast models predict monthly flows and are updated on a weekly basis as new data becomes available as input into the models.
How to access the nowcast models
The end product of the nowcast models are the heatmaps. You can access these heatmaps on our app [here] , where they are updated weekly, as well as in our nowcasting emails, which we send out at the end of the month.
We also have the complete research documentation and code for the nowcast models in our Jupyter Notebooks. You can find the links for the notebooks on our app, and through links provided in the nowcasting emails.
How to interpret our nowcast heatmaps
We have two types of heatmaps:
1) positive/negative heatmap and
2) historical percentile heatmap.
The first heatmap shows the monthly debt flows for GBI-EM countries and the total debt
flows in billions USD.
The intensity of positive and negative flows is based on historical data since 2015. A dark green indicates a historical inflow for that country, while a dark red indicates a historical outflow.
Cells with nowcasted values are indicated by “**”, whereas cells with just numbers show real-time data. Cells where no predicted or real-time data is available are shown as grey cells with “-“ at the center. For those cells, there are nowcast models still in development, out of which Romania, Russia and Brazil will be added soon.
The second heatmap shows monthly debt flows in percentile rank, related to historical flows since 2015.
A value of 100%, or pure white, implies a flow which is higher than any other observed flow for that country. A 0%, or dark blue, indicates a flow lower than any other observed flow.
The purpose of this heatmap is to compare historical flows, regardless of their positive or negative values.
For example, China’s monthly debt flows have been positive since 2015. A low positive flow into China would be positive on the first heatmap, but the second heatmap would indicate a historical low flow for China.
Our nowcasting emails are sent out on the last day of the month. Make sure to check them out and give us feedback!
Our data scientists are eager to respond to any questions you may have.