A vendor says its AI beat buy-and-hold by 13.55%. The paper behind it shows the AI lost money.

Trading Central, a Paris-based provider of market research and analytics, sells an MCP server that lets AI agents and chatbots draw on its data. On October 15, 2026 its blog published 'Same Model, Better Data', a write-up of how its research unit, TC Labs, did in a live trading challenge. The post says that adding Trading Central news summaries and technical indicators 'lifted the accuracy of every AI model they tested' and that the top performer, DeepSeek-R1 with Trading Central data, 'outperformed buy-and-hold by 13.55% on TSLA and 15.57% on BTC in a falling market'². It then points readers to the full paper and to the company's MCP product.

That is a vendor describing its own test, so we opened the paper it links, 'TCLabs @ FinMMEval 2026: Systems for Tasks 1, 2 and 3' by Elvys Linhares Pontes and Mohamed Benjannet, in the CEUR Workshop Proceedings (Vol-4283).

The challenge

FinMMEval 2026 is an evaluation lab run as part of CLEF 2026. In Task 3, each AI agent reads the day's news for Tesla (TSLA) and Bitcoin (BTC) and returns a buy, hold or sell call within a strict three-minute limit. Agents are ranked on cumulative return and compared with a buy-and-hold baseline over the same period. The authors tested three open models (GPT-OSS-20B, Qwen2.5-72B-Instruct and DeepSeek-R1-Distill-Llama-70B) in three setups: organizer data only, organizer data plus Trading Central summaries and indicators, and fine-tuned versions.

What the data did for accuracy

The data helped, though modestly. DeepSeek-R1 scored 52.7 F1 on its own and 53.5 with Trading Central data; accuracy moved from 53.8 to 54.3³. GPT-OSS-20B rose from 48.0 to 50.75 F1 and Qwen2.5-72B from 48.5 to 51.2³. The fine-tuned versions scored slightly below their augmented counterparts, which the authors attribute to constrained reasoning capacity. The paper does not report confidence intervals or the size of the test set, so we cannot say whether gains of 0.8 to 2.7 points are outside noise.

What the leaderboard says

The 13.55% and 15.57% figures are real, but they measure the gap to a baseline that fell hard. On the TSLA leaderboard, buy-and-hold returned -14.82%; on BTC it returned -25.13%⁴. The agent, called TCAgent, returned -1.27% on TSLA and -9.56% on BTC⁴. Subtracting gives the headline numbers: it lost less than holding. It did not make money.

The paper is plain about this. It says the returns 'did not place the proposed approach among the top-performing submissions on the public leaderboard' and that the absolute returns 'remained negative'. The same tables show why. Every listed agent beat the baseline in that window, so beating it was the minimum, not a distinction.

Caveats the authors disclose

The leaderboard is a snapshot from July 3, 2026, and the competition was still running when the paper was written, so final standings may differ. The authors also report that the data feed for technical indicators failed for the BTC task after the evaluation had begun, so the agent ran on less information there than in development. That may explain part of the weaker BTC result, and it also means the BTC number is not a clean test of the product the blog describes. The run uses a single agent per asset, and the paper does not disclose how many trading days the window covers.

How to read a claim like this

Three checks catch most of the problem. First, ask what the comparison is against: beating a baseline that lost a quarter of its value is a different statement from earning a return. Second, ask where the agent ranked among peers that faced the same market. Third, ask who benefits from the framing. The post closes with a pitch for Trading Central's MCP server, and the data in the test came from the same company. None of that makes the study wrong. It is a reason to read the tables before the headline.

The accuracy result is the more useful finding. Giving a model current, structured market context made all three models slightly better at a news-based decision task, which fits a pattern seen across other published tests: what a model reads often matters as much as which model it is. A small, consistent lift on a benchmark is still a long way from a profitable trading agent, and the paper never claims otherwise.