Habr AI suggested using news sentiment as a trading signal for the crypto market

A new breakdown describes an approach where the trading signal for the crypto market is built not on price charts but on changes in news sentiment. The author argues that intraday reversals are more often explained by cascades of publications and reposts than by price history.

Why Signals Break

The article examines a typical problem with technical analysis: moving averages and other indicators adjust to past data and assume that the market regime is sufficiently stable. But in the crypto market, sentiment can switch from bullish to bearish and back several times a day. When this happens, even a carefully tuned strategy begins to produce noise instead of direction because it responds to yesterday's market structure rather than to the fresh stream of messages that is already changing participant behavior.

"News sentiment determines the regime.

The indicator works within the regime."

According to the author, the problem lies not only in the news itself but also in the mechanics of its distribution. Prices are driven not by dry regulatory documents or the publication of a report itself, but by a chain of reposts, comments, and interpretations on the domains and blogs that retail investors follow. Because of this, the market reacts to the media environment faster than classical indicators can adjust, and attempting to apply the same set of signals to all regimes often results in something close to 50/50.

How to Find Sentiment

The practical part is built around vector search across news and posts with subsequent interpretation using an LLM. The author proposes not mixing these tasks: embeddings and cosine distance are needed to extract the relevant news array by meaning, while the language model should separately evaluate the overall sentiment of the sample. Working tools mentioned include combinations using Scrapy, PostgreSQL with PgVector, MongoDB Atlas Vector Search, and for quick starts, Tavily and Perplexity Search API.

• Search not only for exact matches but for semantic context around the asset.
• Don't take the maximum score as the best signal; instead look at borderline mentions.
• First filter influential domains and authors, then refine the query.
• Maintain a strict time window so morning positivity doesn't mix with evening negativity.
• Separate publication search from sentiment analysis so the LLM doesn't substitute interpretation for fact.

Special emphasis is placed on publication time. The article recommends excluding materials without a precise timestamp; otherwise, look-ahead bias enters the selection and the strategy begins to see the future in hindsight. For search, it's proposed to take a two-day range, then filter the last 24 hours on your end. The author considers exactly 24 hours the optimal window: it already provides context but doesn't blur the directedness of sentiment into statistical noise.

Market Testing

To test the hypothesis, the author shows two Bitcoin cases: in one, the query gave neutral-to-bearish sentiment, in another — bullish. These assessments are then compared to the market's subsequent reaction. Formally, this doesn't look like complete academic research, but as an engineering demonstration, the idea reads clearly: news background can be used not as an absolute price predictor but as a regime filter within which entry decisions are made.

The approach then progresses to automation. By the description, an AI agent receives a news signal, opens a position, and holds it until sentiment exhaustion, using a statistically unreachable hard stop and trailing take-profit for risk control. The author also names the weak point directly: if you exit only on sentiment changes, you can lose part of your profits due to parsing delays and news delivery latency.

Therefore, as an improvement, exiting on a 3% pullback from the maximum PnL of the open position is proposed.

What It Means

For algorithmic trading in news-driven and fast-changing markets like crypto, this is another signal that the battle is not only about model quality but also about source quality, timing, and noise filtering. If the author's hypothesis holds up to broader backtesting, priority may shift from complicating indicators to systems that better read market sentiment in real time.