Radar Comex
Methodology, models and artificial intelligence
Where every figure comes from, how it is calculated, and what AI does (and does not do).
Sources
DIAN (customs)
Every import and export declaration since 2001, one by one; customs clearance-time study.
RUES
Business registry: legal name, status, city and activity of each company.
Websites and ProColombia
Contact details published by the companies themselves and the exporters directory.
World Bank and IMF
Logistics Performance Index (LPI) and vessel activity at ports and chokepoints (PortWatch).
Ministry of Transport
Reference freight rates (SICE-TAC) and road cargo manifests (RNDC).
DANE and OpenStreetMap
Municipality codes and coordinates (DIVIPOLA) and the real road network for routing.
United Nations (UN Comtrade)
Monthly trade of other countries by HS code since 2020, with the world and with each country in the Americas; official statistics, no companies.
Brazil: Comex Stat (MDIC)
Brazil's monthly trade with every country by HS code since 2020 and, with Colombia, by mode, customs office and state, with freight, insurance and CIF; open data (Decree 8,777/2016).
Mexico: INEGI
Mexico's merchandise trade by HS code and country (annual), quarterly exports by sector (Indicators Bank) and business directory (DENUE); INEGI free-use terms.
Chile: National Customs Service
Import declarations since 2020, item by item, with CIF, freight, weight and duties (datos.gob.cl, ODC-By licence).
Argentina: ARCA
Imports by declaration item with duties paid, and exports by product and country (public-access information).
United States: U.S. Census Bureau
Monthly imports and exports by product and country since 2020, with CIF and air and vessel weight (public domain).
Canada: Statistics Canada
Monthly trade by product and country since 2020 (Open Government Licence – Canada), converted to US dollars at the Bank of Canada rate.
Paraguay: National Customs Directorate
Monthly customs declarations since 2020 with FOB, freight, insurance and weight (open data portal).
Superintendency of Companies
Financial statements reported by companies.
World Trade Organization
Tariffs of Colombia and destination countries, and technical-measure notifications (ePing).
Financial Superintendency
Market exchange rate (TRM), published on datos.gov.co.
All of them are public sources. Every screen in the platform states where its data comes from.
The rules we follow
No figure without backing
Every number comes from a query on the data; if it could not be measured, we say so.
Limits next to the figure
An incomplete year or a small sample is flagged next to the figure, not in fine print.
Medians, not averages
A single outlier declaration does not move prices or times: we use medians and percentiles.
The period is the file's
Each declaration counts in the month DIAN published it, and each year is checked against its control totals.
Same period vs. same period
A year in progress is compared with the same months of the previous year, never with a full year.
Correlation is not causation
A figure changing after a decree does not prove the decree changed it.
Statistical and data-science models
We run models on the same data, recalculated on the loaded data. Each one is validated on data it did not see and compared against a simple alternative; if it does not beat it, we say so. Today our team uses them to prepare analyses and custom reports; the figures below come from their latest run.
Six-month forecast
How much of each HS code will be imported and exported over the next six months, with an uncertainty band.
Method
Seasonal Holt-Winters, always compared against a naive forecast on months the model did not see.
Measured validation
4,154 import series forecast; in 2,087 the model erred less than the naive one, and the naive forecast is used for the rest.
Does the exchange rate improve the forecast?
Whether adding past exchange-rate changes reduces forecast error.
Method
Forecast with and without lagged exchange rate, against a placebo with a shuffled rate.
Measured validation
It helped in 1,213 of 3,831 import series, versus 986 with the placebo: a small gain, reported as is.
Price outliers
Which transactions were priced far above or below normal for their product and origin.
Method
Robust score with median and MAD by HS code and country; an outlier if |z| ≥ 3.5 and also 1.5 times above or below.
Measured validation
6,976 outlier import transactions in 2025, across 1,815 HS codes.
Importer and exporter churn
Which companies are most likely not to trade next year.
Method
Life table by company profile, shrunk towards the rate for its size; validated on a year it did not see.
Measured validation
Discrimination (AUC) of 0.84 for importers, versus 0.80 using size alone.
Market recommender
Which countries an exporter could sell to, based on where others with similar products sell.
Method
Countries scored by the company's product mix; validated against the countries they actually opened later.
Measured validation
Of 3,263 exporters that opened a new country in 2025, for 52 % that country was among their top 5 suggestions, versus 42 % suggesting the most popular destinations.
HS code classifier
Under which HS code importers usually declare goods described in words.
Method
Naive Bayes on declaration descriptions; it does not replace DIAN's official classification.
Measured validation
The right HS code is among the top three suggestions in 26 % of cases (chance: 1.9 %). It helps you get oriented; it is not an answer.
Supplier dependence risk
Which importers rely on a single supplier or country for their products.
Method
Herfindahl-Hirschman concentration index by supplier and country over the last 18 months.
Measured validation
18,453 importers assessed; 9,229 with high concentration.
Data-load monitoring
Whether a loaded month has figures that do not fit the months around it.
Method
Each month against the median of its twelve neighbours, across seven measures; outliers are reviewed by a person.
Measured validation
616 months checked; 26 flags to review.
Prospect scoring
Which importers show the most signs they could benefit from the platform.
Method
Scoring rule with declared weights (not a trained model: there is no customer history to train it on yet).
Measured validation
29,531 importers scored with 2025 data.
Use of artificial intelligence
We use generative AI —currently Claude, by Anthropic— for specific tasks:
- Writing plain-language summaries of an HS code's figures. Every number in the summary is checked against the data; if it does not match, the summary is discarded.
- Explaining foreign-trade regulations: what changes, whom it affects and from when. These are shown as unconfirmed until a person reviews them.
- Suggesting a company's website, which is then verified independently before it is saved.
- Classifying, as part of security monitoring, incoming texts that show signs of an attack.
Figures are never produced by AI: they come from queries on the data and the models described above. No decision about your account is made by AI alone, and your user data is not used to generate summaries or explanations.