Social Media Sentiment Signal With Bot Filtering

When to Use

Use this skill when a strategy consumes social-media text for a ticker and the feed has to be cleaned before it is aggregated. Stock microblog streams are contaminated at scale: Cresci et al. (2019) analysed 9M stock-related tweets across the five main US markets and found coordinated bot groups piggybacking on the cashtags of liquid names to promote low-value stocks, with as much as 71% of the authors of suspicious financial tweets classified as bots. A sentiment mean taken over that stream measures a campaign, not an opinion.

The engine screens in two places, because the two kinds of abuse are visible in different places:

Per post — account age, posting-rate burst, spam-pattern match. These are decidable from one post and live in filter_post.
Per batch — near-duplicate text collapse and one-vote-per-author aggregation. Coordination is a property of the group; no per-post rule can see it. These live in process_social_posts.

It then scores the survivors with a financial lexicon and standardizes against a baseline — or refuses to, when too few independent contributors remain.

When NOT to Use

As a bot detector. These are metadata hygiene rules, not detection. Cresci et al. (2017) benchmarked Twitter itself, human annotators and state-of-the-art tools against modern social spambots and found none of them accurate (human annotators scored under 24%). Assume a competent operator passes every rule here.
On a feed whose user counters were backfilled. user_account_age_days, user_follower_count and user_posts_last_hour are as-of-query values. Fetched today for a two-year-old post they describe the account now, so every historical account passes the age screen in hindsight and the backtest is measuring a filter that never ran. These fields must be snapshotted at ingestion.
As a standalone entry signal. The output is one alternative-data input. Sentiment Z-scores do not carry position sizing, stops or exposure limits — see liquidity-adjusted-position-sizing and kill-switch-and-drawdown-circuit-breakers.
With a baseline computed on individual post scores. See Prerequisites — the units must match, or every signal is muted by roughly $\sqrt{n}$.
Where the data licence does not permit it. Platform terms govern what may be stored, redistributed and traded on, independently of what the API returns. Confirm entitlement before building on a feed — alternative-data-vendor-due-diligence-checklist, data-vendor-contractual-usage-restriction-tracking.
To generate the posts as well as read them. Posting to move a price you trade is market manipulation. This skill reads a public stream; it must never be wired to one that writes.

Prerequisites

SocialPost records with post_id, asset_id, user_id, created_at_iso, text, user_account_age_days, user_follower_count, user_posts_last_hour, is_verified_user.
Timezone-aware ISO-8601 timestamps. Naive strings are rejected. X's API v2 stamps created_at in RFC 3339 UTC with a trailing Z, so live feed data already complies; a bare 2026-08-05 does not and is refused rather than assumed to be UTC.
A baseline in the right units. historical_baseline_mean and historical_baseline_std must describe the distribution of this same filtered aggregate statistic — e.g. past daily filtered means for this asset under the same filters and window. They must not be the mean and standard deviation of individual post scores: the standard deviation of a mean of $n$ observations is smaller than the per-post standard deviation by roughly $\sqrt{n}$, so a per-post $\sigma$ in the denominator understates every $Z$ by that factor. The engine cannot detect which one it was handed.
Thresholds you are willing to defend: min_account_age_days, max_posts_per_hour, min_effective_sample, signal_z, strong_signal_z. The defaults are house heuristics, not standards — see references/standards.md.

Workflow

Validate and align the batch:
- Every post is validated before anything is aggregated; a malformed post fails the batch rather than being dropped silently.
- Decision point — a post whose asset_id differs from the requested asset raises. Aggregating another ticker's posts under this label would publish a signal for an instrument it was never measured on. Filtering them out quietly would hide a broken ingestion join.
Apply the point-in-time cutoff:
- With as_of, posts stamped after it are excluded and counted in future_posts_excluded_count; with lookback_window_minutes set, posts older than the window are counted in stale_posts_excluded_count. The retained window is inclusive at both ends.
- Decision point — without as_of no cutoff is applied at all. The engine does not guess one. A backtest that omits it is reading the future and the counters will read zero, which is not evidence of correctness.
- Configuring lookback_window_minutes without passing as_of raises rather than silently ignoring the window.
Run the per-post screens (filter_post):
- Unverified accounts below min_account_age_days; accounts strictly above max_posts_per_hour; the first matching spam pattern.
- Decision point — a verification badge buys no exemption by default. X's published criteria for the blue checkmark are a paid Premium subscription plus a display name, photo, confirmed phone and recent activity; X states it does not mean the account was ID verified. On 5 December 2025 the European Commission fined X €120M under the DSA, finding the checkmark deceptive because "anyone can pay to obtain the 'verified' status without the company meaningfully verifying who is behind the account". Set trust_verified_accounts=True only for a platform whose badge you have independently established to mean identity.
Run the batch-level coordination screens (process_social_posts):
- Near-duplicate collapse: post text is normalised (links, cashtags, mentions, punctuation stripped) and repeats are dropped, keeping the first.
- One vote per author: each author's posts are averaged to a single contribution before the cross-author mean, so a prolific account cannot carry the batch.
- Decision point — the effective sample size is the contributor count, not the post count. Thirty identical posts from thirty aged, low-rate accounts is one contribution, not thirty.
Score sentiment and gate the signal:
- Each distinct lexicon term counts once per post; a term within three tokens after a syntactic negator has its polarity flipped, so "not bullish" scores $-1.0$.
- Decision point — below min_effective_sample contributors there is no Z-score at all. sentiment_z_score is None and directional_signal is INSUFFICIENT_DATA. None means "not measurable" and must be rendered as such, never coerced to 0.
- Decision point — INSUFFICIENT_DATA is not NEUTRAL. NEUTRAL means the balance of opinion was measured and was flat. A consumer may reasonably act on the second and must never act on the first.
Standardize and band:
- $Z = (\mu_{\text{filtered}} - \mu_{\text{baseline}}) / \sigma_{\text{baseline}}$, with $\sigma_{\text{baseline}}$ validated finite and strictly positive at construction.
- Decision point — the band is decided on the unrounded $Z$. Rounding first promotes $1.9951$ to $2.0$ and hands back a STRONG signal the data does not support. The reported figure is rounded; the decision is not.

Full procedure: see references/workflows.md. Standards reference: see references/standards.md. Printable pre-flight checklist: see assets/checklist.md.

Common Pitfalls

Trading a signal built from one post: a single surviving post scoring $+1.0$ against a $\sigma$ of $0.15$ yields $Z \approx 6.3$ — a maximum-conviction signal off a sample of one. The sample gate exists for this; do not disable it by setting min_effective_sample=1 to make a thin ticker "work".
Screening accounts but not campaigns: thirty established, low-rate accounts posting the same line pass every per-post rule. Per-account heuristics do not see coordination; the duplicate-text and per-author screens do.
Treating a paid checkmark as identity: exempting "verified" accounts from the age screen re-opens the cheapest bypass there is — a €3/month subscription.
Backfilled account metadata: querying account age today for historical posts makes every account look established and produces a backtest of a filter that never fired.
A per-post baseline in the denominator: mixing the units mutes every signal by roughly $\sqrt{n}$, and the output still looks like a valid Z-score.
Absorbing a degenerate baseline into $Z=0$: a zero, negative or NaN $\sigma$ has no Z-score interpretation. The engine raises at construction rather than reporting a confident NEUTRAL built on a broken config.
Rounding before banding: round(1.9951, 2) == 2.0 silently upgrades BULLISH to STRONG_BULLISH.
Scoring pump as bullish: it is the vocabulary of the manipulation this skill screens for. A pump post that clears the filters should not also push the signal the way the campaign intends. moon, rocket and breakout carry the same risk and are kept only because they are also ordinary retail vernacular.
Reading call/put as direction: they are instrument names. "Sold calls" is a bearish position scored bullish here — strip or reweight them on an options-heavy venue.
Assuming the spam regexes hold: t(dot)me and Unicode period look-alikes are normalised, but obfuscation is adaptive and a rewritten payload defeats the list the same day.
Aggregating another ticker's posts: a broken ingestion join publishes a signal for an instrument it was never measured on. The engine raises on an asset_id mismatch.

Verification

Instantiate SocialMediaSentimentSignalWithBotFilteringEngine(historical_baseline_mean=0.0, historical_baseline_std=0.1, min_effective_sample=20). Feed 12 authors posting distinct bullish text and 8 posting distinct bearish text: mean $= (12-8)/20 = +0.2$, $Z = 0.2/0.1 = 2.0$ exactly, banded STRONG_BULLISH at the inclusive boundary.
Feed 30 established, low-rate accounts posting the same line: confirm bot_posts_filtered_count == 0, duplicate_posts_filtered_count == 29, effective_sample_size == 1, and INSUFFICIENT_DATA with sentiment_z_score is None.
Feed a single clean bullish post: confirm INSUFFICIENT_DATA, not STRONG_BULLISH.
Screens: a 5-day-old account posting 100/hr with a t.me/ link yields exactly three rejection reasons; a 1-day-old verified account is still rejected unless trust_verified_accounts=True; age 29 rejects and 30 passes; 41 posts/hr rejects and 40 passes; t (dot) me/, t dot me/ and t․me/ all match.
Scoring: "not bullish" $= -1.0$; "moon moon moon moon" $=$ "moon"; "pump it" $= 0.0$.
Point-in-time: posts stamped after as_of land in future_posts_excluded_count; 2026-08-05T08:30:00-04:00 is after a 12:00Z cutoff; a naive as_of, a naive post timestamp, and a configured lookback_window_minutes with no as_of all raise.
Banding: with historical_baseline_mean=-1.9951, historical_baseline_std=1.0 and a balanced batch, $Z = 1.9951$ must band BULLISH, not STRONG_BULLISH.
Negative checks: historical_baseline_std of $0$, negative, NaN or infinity; min_effective_sample=0; signal_z >= strong_signal_z; an empty or non-string text; a negative counter; an asset_id mismatch — all raise ValueError.
Run python -m unittest discover -s skills/social-media-sentiment-signal-with-bot-filtering/scripts and confirm a 100% pass rate.

Social Media Sentiment Signal With Bot Filtering

When to Use

When NOT to Use

Prerequisites

Workflow

Common Pitfalls

Verification

Hands off to 9

Handed off from 3

When to Use

When NOT to Use

Prerequisites

Workflow

Common Pitfalls

Verification

Related Skills

Hands off to 9

Handed off from 3