Method and limits
The sentiment index counts shipping market news by machine and balances the items that lean bullish against those that lean bearish for freight earnings. This page says which news is used, how it is labelled, how often the labels agree with a person, and what the index cannot tell you.
Method fixed on 2026-10-19; label check 2026-10-19.
Sources
Headlines come from the RSS feeds of Splash247, All About Shipping, Marine Log (from January 2024), and since 2026-10-19 also Offshore Energy and The Maritime Executive. A source is used when its feed is public, its robots.txt allows our reader, and its terms do not forbid automated reading or reuse. We store what the feed gives (headline, short summary, link, date) and show only the headline, the link and our labels; never the article text. Publishers who prefer not to be listed can write to contact@vesselvector.com.
The index uses only the three sources with history, for every month, so that adding sources does not move it.
Labels
Each headline with its feed summary is labelled by TypeSafe's Jev model (version jev-1.13.0, fixed so that labels do not change silently): which segment it is about, whether it concerns merchant shipping markets, its topic, and whether it points to higher or lower freight earnings, with a confidence. The rules given to the model treat newbuilding orders, company results and deals, and regulation without a stated rate effect as neutral. The same news reported by several sources within a few days counts once.
Which items count
- Topics: freight rates, demand and trade, geopolitics and routes, ports and canals. Company, newbuilding, ship-sale and regulation news is left out (see accuracy).
- Only items labelled bullish or bearish with a confidence of at least 0.6; from 0.8 they count in full, below that at 0.5.
- Index for a month = (weighted bullish − weighted bearish) ÷ (weighted total), by publication date. It runs from −1 to +1.
- All segments together: per segment there are too few items a month (dry bulk often fewer than five).
- The index leans bullish: it was above zero in 31 of 33 months. Most counted items are about geopolitics and routes, and there attacks and detours (603 bullish items since January 2024) far outnumber their easing (48 bearish). Compare months with each other rather than reading the level on its own.
How accurate the labels are
Randomly drawn items were labelled by hand, by one person at VesselVector, with a second person re-checking 40 of them, without seeing the model's answer, and compared with the model on items where it gave a direction on a market topic:
| Model answer | Items | Agree | 95% range | Opposite direction |
|---|---|---|---|---|
| Confidence 0.8 or more (full weight) | 32 | 91% | 76%–97% | 0 |
| Confidence 0.6 to 0.8 (half weight) | 38 | 71% | 55%–83% | 0 |
When the model disagreed it was almost always calling a neutral item bullish or bearish, not getting the direction the wrong way round. On company, newbuilding, ship-sale and regulation news its directions agreed only 4 times in 20, which is why those topics are left out. The samples are small; the ranges show how far the true rate could be.
What the index does not show
We compared the monthly index from 2024-01 to 2025-12 with two public series, the month the news appeared and up to three months later, and kept 2026-01 onwards aside as a check:
- US Bureau of Labor Statistics producer price index, deep sea freight transportation (PCU483111483111), monthly
- IMF PortWatch world seaborne import volume of container, dry bulk and tanker cargo, monthly
None of the 8 comparisons showed a link (correlations of 0.18 or less, p-values of 0.39 or more). With this much data only a strong link, a correlation of about 0.54 or more, would have shown; a weaker one cannot be ruled out. Both series are rough stand-ins: segment freight indices are licensed and were not used. So the index describes the tone of the news; it is not a guide to future rates.
Other limits
- English-language trade press only, and headlines rather than full articles.
- A month with few items moves a lot on a single story; the item count is shown next to every value.
- One person labelled the items and a second re-checked 40 of them; more checks will be added over time and the figures above updated with their date.