Habr AI piece proposes using news sentiment as a trading signal for crypto markets

A new write-up describes an approach in which the trading signal for the crypto market is built not on charts but on changes in news sentiment. The author claims that intraday reversals are more often explained by a cascade of publications and reposts than by price history.

Why signals break

The article examines a typical problem of technical analysis: moving averages and other indicators adapt to past data and assume that the market regime is reasonably stable. But in crypto, sentiment can flip from bullish to bearish and back several times a day. When that happens, even a carefully tuned strategy starts producing noise instead of direction, because it reacts to yesterday's market structure rather than to the fresh stream of messages that is already changing participants' behavior.

"News sentiment determines the regime.

The indicator works inside the regime."

According to the author, the problem lies not only in the news itself but in the mechanics of how it spreads. Price is moved not by dry regulatory documents or the publication of a report itself, but by the chain of reposts, comments and interpretations on the domains and blogs that retail investors follow. As a result, the market reacts to the media environment faster than classic indicators can readjust, and applying the same set of signals to every regime often ends in a result of around 50/50.

How to find sentiment

The practical part is built around vector search over news and posts, followed by interpretation with an LLM. The author recommends not mixing these tasks: embeddings and cosine distance are for retrieving a semantically relevant body of news, while the language model should separately assess the overall mood of the sample. The tools mentioned include setups built on Scrapy, PostgreSQL with PgVector, and MongoDB Atlas Vector Search, plus Tavily and the Perplexity Search API for a quick start.

• Search not only for exact matches but for semantic context around the asset.
• Do not take the maximum score as the best signal; look at borderline mentions.
• First select influential domains and authors, and only then refine the query.
• Keep a strict time window so that morning positivity does not mix with evening negativity.
• Separate retrieving publications from analyzing sentiment, so the LLM does not substitute its own interpretation for fact.

The article puts particular emphasis on publication time. It advises excluding material without an exact timestamp, otherwise look-ahead bias creeps into the sample and the strategy starts seeing the future in hindsight. For retrieval, it suggests requesting a two-day range and then filtering the last 24 hours on your own side. The author considers a day the optimal window: it already provides context but does not yet dilute the direction of sentiment into statistical noise.

Testing against the market

To test the hypothesis, the author shows two Bitcoin cases: in one the query returned neutral-bearish sentiment, in the other bullish. These readings are then compared with the market's subsequent reaction. Formally, this does not look like a full academic study, but as an engineering demonstration the idea comes across clearly: the news backdrop can be used not as an absolute price predictor but as a regime filter, within which the decision to enter is then made.

The approach is then taken to automation. As described, an AI agent receives the news signal, opens a position and holds it until sentiment is exhausted, using a statistically unreachable hard stop and a trailing take-profit to control risk. The author also names the weak spot directly: if you exit only on a change in sentiment, you can lose part of the profit because of delays in parsing and delivering news. As an improvement, the article proposes exiting on a 3% pullback from the open position's maximum PnL.

What this means

For algorithmic trading in news-driven, fast-moving markets like crypto, this is one more sign that the contest is being fought not only over model quality but also over source quality, timing and noise filtering. If the author's hypothesis holds up in a broader backtest, priority may shift from building more complex indicators to systems that better read market sentiment in real time.