Methodology

How tracking works

XAlphaTrack monitors public tweets from financial analysts on X continuously. twice daily, a cron worker fetches new tweets using the X API (GetXAPI service). Each tweet is stored in full, including the text, timestamp, engagement counts, and any quoted tweets. This data is collected from January 2026 onwards.

How calls are detected

Once tweets are ingested, they are classified by a large language model. The classifier reads each tweet and extracts whether it is an investment call, and if so, what the ticker(s), direction (long/short/exit), conviction level, horizon, and reasoning are.

Key rules applied:

• One row per ticker: If a tweet mentions multiple tickers, it creates multiple call rows (one per ticker) — all rows share the same direction and conviction.
• Named-entity calls: Calls where the analyst names a company without a ticker symbol are classified as calls with an entity name instead of a ticker. These are kept private until a ticker resolution pass assigns them a symbol.
• Confidence score: Every classified tweet receives a confidence score (0–1) reflecting the classifier's certainty that the call is a genuine investment call.
• Not-a-call verdicts: Tweets that mention tickers but are not investment calls (e.g. news sharing, sarcasm, discussion) are marked with a classification_note and do not create call rows.

How scoring works

Each call is scored by comparing the ticker's price movement to expected horizons:

• Entry point: The first closing price at or after the tweet's timestamp is used as the entry.
• Measurement horizons: Price moves are measured at 1-day, 1-week, 1-month, 3-month, and 6-month intervals after the call.
• Benchmark: Each call is compared to the S&P 500 (SPY) over the same period to control for market conditions.
• Non-US tickers: Tickers not trading in US exchanges are marked with "—" and do not count as misses in hit-rate calculations — data is insufficient to score them accurately.
• Minimum call threshold: Analyst rankings on the leaderboard only count calls where the analyst has at least 10 scored calls (calls old enough to have a 1-month return).

Known limitations

• AI classification can be wrong: The classifier is not perfect. Some calls are misclassified (false positives) and some are missed (false negatives). Every public call detail page displays the classification confidence so you can assess the certainty of each extracted call.
• Per-tweet direction: All tickers mentioned in the same tweet share the call direction. If an analyst tweets "Long $NVDA, short $AMD", both calls are recorded with their respective directions — but this is a known point of improvement (per-ticker directions coming later).
• Engagement counts are snapshots: Like, retweet, and reply counts are recorded when the tweet is captured, not updated in real time.
• Price data availability: Non-US stock prices are not yet fetched. Crypto tickers may have limited availability. These calls are marked with incomplete pricing status.

Have a question or found an issue?

If you spot a misclassified call, use the "Report an issue with this call" button on any call detail page. If you have other questions about the methodology or data, you can reach us by replying to any email we send you.