On a Saturday morning in late October, 30 of the newest 60 posts on CVS carried a Bearish tag and none carried a Bullish one. Read naively, that is a stock the crowd has turned on. It was one account: 29 of the 30 tags came from the same user, mostly reshares sent within a few hours of each other.
That made me ask a plainer question first. Stocktwits lets every author mark a post Bullish or Bearish, and those tags are what most "Stocktwits sentiment" projects count. How many posts actually carry one, and how much should a ratio built from them be trusted? I pulled recent posts for 15 tickers in Python and counted. The short answer: about one post in three is tagged, and a tag rate tells you as much about who is posting as about what they think.
The data
I used the Stocktwits API, which returns Stocktwits streams and sentiment readings as JSON. Three endpoints did the work. /symbol-stream returns a symbol's posts newest first, 30 per page, and I paged back once with cursor.max to get the newest 60 per symbol. /sentiment returns Stocktwits' own model reading for the symbol. /trending-symbols gave me the trending list I picked part of the sample from.
The 15 symbols are a deliberate mix: index ETFs (SPY, QQQ), the most discussed large caps (AAPL, TSLA, NVDA, AMZN, AMD), retail favourites (PLTR, GME), bitcoin (BTC.X) and bitcoin-adjacent names (MSTR, CLSK), plus CVS, V and KO from the US trending list that morning (CLSK, MSTR and AMZN were on it too). I collected everything on Saturday 24 October 2026, around 11:00 UTC, with the US market closed. That is 900 stream rows and 801 unique posts, from 56 API calls.
Each post carries an id, created_at, a symbols list (every cashtag in the post) and entities.sentiment, which is {"basic": "Bullish"}, {"basic": "Bearish"} or null when the author did not tag it. Posts also carry the author's profile. I dropped all of it before saving anything: no usernames, ids, names, avatars or message text are stored, and nothing below quotes an individual post.
The code
To run it, set RAPIDAPI_KEY to a key from the Stocktwits API listing on RapidAPI, which includes a free plan, and install requests and pandas. Responses are saved to a local folder, and each post is cut down to the four fields used here before it is written, so nothing about the authors lands on disk. It spends 45 calls once (two stream pages and one sentiment reading per symbol) and nothing after that.
import json
import os
from pathlib import Path
import pandas as pd
import requests
HOST = "stocktwits5.p.rapidapi.com"
CACHE = Path("stocktwits_cache")
SYMBOLS = ["AAPL", "TSLA", "NVDA", "AMZN", "AMD", "PLTR", "GME", "SPY", "QQQ",
"BTC.X", "CLSK", "CVS", "MSTR", "V", "KO"]
def slim(post):
"""Keep what the analysis needs and nothing about the author."""
return {"id": post["id"], "created_at": post["created_at"],
"symbols": [s["symbol"] for s in post.get("symbols") or []],
"sentiment": ((post.get("entities") or {}).get("sentiment") or {}).get("basic")}
def get(path, **params):
"""One API call, cached on disk so a re-run costs nothing."""
name = path.strip("/") + "_" + "_".join(f"{k}-{v}" for k, v in sorted(params.items()))
file = CACHE / f"{name}.json"
if file.exists():
return json.loads(file.read_text())
r = requests.get(f"https://{HOST}{path}", params=params, timeout=60,
headers={"x-rapidapi-host": HOST,
"x-rapidapi-key": os.environ["RAPIDAPI_KEY"]})
r.raise_for_status()
data = r.json()
if path == "/symbol-stream":
data["data"]["messages"] = [slim(m) for m in data["data"]["messages"]]
CACHE.mkdir(exist_ok=True)
file.write_text(json.dumps(data))
return data
rows, scores = [], {}
for sym in SYMBOLS:
params = {"symbol": sym}
for page in range(2): # 2 pages x 30 = the newest 60 posts
data = get("/symbol-stream", **params)["data"]
rows += [{"stream": sym, **m} for m in data["messages"]]
params = {"symbol": sym, "max": str(data["cursor"]["max"])}
now = get("/sentiment", symbol=sym)["data"]["data"]["sentiment"]["now"]
scores[sym] = {"model_score": now["value"], "label": now["label"],
"label_normalized": now["labelNormalized"]}
df = pd.DataFrame(rows)
df["tagged"] = df["sentiment"].notna()
df["bullish"] = df["sentiment"].eq("Bullish")
per = df.groupby("stream").agg(posts=("id", "size"), tagged=("tagged", "sum"),
bullish=("bullish", "sum"))
per["tagged_pct"] = (per["tagged"] / per["posts"] * 100).round(1)
per["bullish_of_tagged_pct"] = (per["bullish"] / per["tagged"] * 100).round(1)
per = per.join(pd.DataFrame(scores).T)
print(per.sort_values("tagged_pct", ascending=False).to_string())
uniq = df.drop_duplicates("id") # a post naming $TSLA and $SPY sits in both streams
tagged = uniq[uniq["tagged"]]
multi = tagged["symbols"].map(len).gt(1).sum()
print(f"{len(df)} stream rows, {len(uniq)} unique posts")
print(f"tagged: {len(tagged)} ({len(tagged) / len(uniq):.1%}), "
f"bullish share of tags: {tagged['bullish'].mean():.1%}")
print(f"tagged posts naming more than one ticker: {multi} ({multi / len(tagged):.1%})")
Two details in there matter. Paging uses cursor.max from the previous response, which the API documents as "messages posted before it". And the totals are computed on drop_duplicates("id"), because the same post shows up in every stream whose cashtag it names. Of my 900 rows, 99 were repeats.
Results
Of the 801 unique posts, 286 carried a tag: 35.7%. The rest, almost two thirds, say nothing machine-readable about direction. Per symbol, the tag rate runs from 55% on BTC.X down to 18% on SPY and V. The busiest retail names sit near the top (PLTR 48%, AMD and GME 45%), the index ETFs near the bottom (QQQ 25%, SPY 18%). Keep the sample size in mind: one symbol is 60 posts, so a handful of posts moves its rate by ten points. Treat the order as rough.
There is no obvious reason in the data why some posts are tagged and others not, but two patterns are clear. Replies are almost never tagged: 3 of 37 replies (8.1%) against 37.0% of original posts. And posts with an image or GIF are tagged more often, 49.4% of 237, although that number is inflated by the CVS account below, which tagged GIF reshares.
Among the tagged posts, Bullish wins: 178 against 108, or 62.2%. It leans Bullish on 11 of the 15 symbols, most strongly on AMZN (21 of 22 tags) and CLSK (19 of 21). The four that lean Bearish are SPY (5 of 11 Bullish), QQQ (5 of 15), MSTR (7 of 22) and CVS (0 of 30). With 11 to 33 tags per symbol, these are directions, not measurements.
How does this compare with Stocktwits' own reading? /sentiment returns a value that the API documentation describes as a model score from -1 (bearish) to +1 (bullish). Across the 15 symbols, the Spearman rank correlation between my Bullish share of tags and that score is 0.87, or 0.86 without CVS. So the two broadly agree on which symbols lean which way. I would not read more into it than that: Stocktwits does not publish what goes into the score, and if author tags feed it, the agreement is partly the same data counted twice. CVS is the useful exception. Its 30 tagged posts are 100% Bearish, while the model scores it -0.08, labelled NEUTRAL, which suggests the model looks at more than these tags or at a longer window.
What this data cannot tell you: whether any of these tags say anything about future prices. I did not test that, and a single Saturday snapshot of 60 posts per symbol could not. This article stops at measuring the signal itself.
What tripped me up in the data
One account can be a whole stream. CVS was number 2 on the trending list that morning, yet its newest 60 posts reached back 184 hours, to 16 October. Most of its tags came in one burst: 22 of the 30 on Monday 19 October, 25 of them reshares, 19 with a GIF or image. To see who was behind them without storing anyone's identity, I re-read the same 60 posts and replaced each author with a throwaway label (a1, a2, ...) that is not derived from the user id. Only 2 accounts tagged anything on CVS, and one of them wrote 29 of the 30 tags.
The other four high-tag streams look like crowds. On BTC.X, 25 different accounts tagged 33 posts and the busiest wrote 5. On PLTR, 26 accounts tagged 29 posts and nobody wrote more than 2. If you build a Bullish ratio from tags, count distinct taggers next to it. In memory, before you throw the author away, it is one line on the raw data["messages"] list of a /symbol-stream response:
from collections import Counter
taggers = Counter(m["user"]["id"] for m in messages if m["entities"]["sentiment"])
One tag, many tickers. The tag belongs to the post, not to a symbol. 101 of the 286 tagged posts (35.3%) name more than one cashtag, some of them five. A post tagged Bullish that names TSLA, SPY and QQQ is counted as Bullish in all three streams, whatever the author meant. Restricting to posts with a single cashtag is the clean fix; it kept 509 of the 801 posts and 185 of the tags.
Reshares do not inherit the tag. A reshare has its own entities.sentiment, and the post it shares sits under reshare_message.message with its own. Of 163 reshares, 69 carry their own tag and 83 share a post that was tagged; in 29 of them only the shared post has a tag. Whether those count is your call, but decide it on purpose, because the default (read the outer post only) silently drops them.
label and labelNormalized disagree more often than they agree. The /sentiment reading has two bands: label for the raw value (the model score), and labelNormalized for valueNormalized (the normalized score), which the documentation describes as the reading on a 0 to 100 scale. They gave the same band on just 1 of my 15 symbols (SPY). On 5 they point in opposite directions, and on 2 more one says extremely bullish while the other says neutral:
| Symbol | Model score | Score band | Normalized score | Normalized band | Bullish share of tags |
|---|---|---|---|---|---|
| GME | 0.98 | EXTREMELY_BULLISH | 34 | BEARISH | 81% |
| BTC.X | 0.78 | EXTREMELY_BULLISH | 28 | BEARISH | 67% |
| AAPL | 0.47 | BULLISH | 31 | BEARISH | 57% |
| TSLA | 0.33 | BULLISH | 36 | BEARISH | 53% |
| V | 0.20 | BULLISH | 43 | BEARISH | 64% |
| AMD | 0.81 | EXTREMELY_BULLISH | 52 | NEUTRAL | 78% |
| PLTR | 0.81 | EXTREMELY_BULLISH | 47 | NEUTRAL | 76% |
A score of 0.98 that normalises to 34 only makes sense if the 0 to 100 scale is relative, most likely to the symbol's own history. The response does not say so, so I cannot confirm it, but the practical rule is clear: label answers "bullish or bearish?", labelNormalized looks like it answers "more or less bullish than usual?", and mixing them in one dashboard gives you GME as both.
Trending is not the same as busy. CVS (trending number 2), V (9) and KO (24) were all on the trending list, but their newest 60 posts covered 184, 90 and 95 hours. SPY, not trending at all, filled 60 posts in 4.6 hours. A fixed number of posts per symbol is a fixed sample size, not a fixed time window, so tag rates for slow symbols mix several days of mood. If you compare symbols, page by time (since and max both take a message id) rather than by count.
Weekend data is different data. All of this was collected on a Saturday, and it shows: 78% of the newest 60 TSLA posts named more than one ticker, and BTC.X, which trades all weekend, filled its 60 posts in 5 hours. A weekday snapshot would give different rates. That is a limit of this sample, not a finding about weekends.
What I would do next
Three things. Collect the same 15 symbols at the same hour every day for a few weeks with since, so each symbol has a time window instead of a post count. Report every tag ratio next to its count of distinct taggers, and flag any stream where one account writes more than a fifth of the tags. And only then ask whether tags lead or lag price, with the /sentiment-daily history and real out-of-sample testing, which is a different and much longer article.
The data came from the Stocktwits API. If you need Stocktwits streams collected on a schedule or delivered as files, get in touch.



Top comments (0)