Social and influencer feeds
This skill is mostly about what you cannot do, and the honest version of that is short: in 2026 there is no free read access to X, Reddit's own help centre says academic research through its Data API is a policy violation, and StockTwits' developer docs return 404. What you can get is a sample that has already been filtered by deletion and by your own attention, and both filters point the same way - toward a backtest that looks better than the live signal.
Everything marked ✅ Measured is printed by scripts/social_feeds.py (numpy + pandas, seed
20260910, under 2 s, no network, no scraper, no API client, no credential path). Its
posts are synthetic and seeded. Everything marked ✅ source-verified was read at the URL
given, on 2026-09-10; 🔴 marks a page that could not be reached from here at all.
1. Access, checked live on 2026-09-10
X / Twitter
🚨 The tier model is gone. ✅ docs.x.com/x-api/getting-started/pricing: "The X API
uses pay-per-usage pricing. No subscriptions-pay only for what you use." There is no
Free, Basic or Pro tier anywhere in the current docs.
| item | value, ✅ read 2026-09-10 |
|---|---|
| Posts: Read | $0.005 per resource, charged "per resource returned in the response" |
| User: Read | $0.010 · Following/Followers: Read $0.010 · Like/Mute/Block: Read $0.001 |
| Owned Reads (your own app's data) | "$0.001 per resource (1,000 resources for $1)." |
| Post: Create | $0.015 per request; $0.200 with a URL |
| Monthly cap | "Pay-per-usage plans are capped at 3 million Post reads per monthly billing cycle." Above that: Enterprise. |
| Deduplication | same resource re-read inside a 24-hour UTC window is free - X calls it "a soft guarantee" |
| Empty balance | "API requests will be blocked until you add credits" |
🔴 Could not be reached: developer.x.com/en/portal/products and
developer.x.com/en/products/x-api return HTTP 402 to a fetch and redirect to a login
wall in a browser; developer.x.com/en/developer-terms/agreement-and-policy likewise
402. ⚠️ So whether legacy Free/Basic/Pro subscriptions still exist for grandfathered
accounts is UNVERIFIED. The quotes below are from docs.x.com/developer-terms/ restricted-use-cases, which is X's current docs host, not from the Agreement itself.
Redistribution, ✅ verbatim:
- "You may only distribute up to a total of 1,500,000 Post IDs to a single entity within a 30 day period" - per 30 days, not per day.
- Hydrated content: "via non-automated means" only, "up to 50,000 hydrated public Post Objects and/or User Objects per recipient, per day", and "should not make this data publicly available (for example, as an attachment to a blog post or in a public Github repository)."
- Academic institutions may redistribute unlimited Post IDs / User IDs for "non-commercial research".
🔑 So a reproducible social dataset is a list of IDs, and every replicator has to re-buy the content. That is a reproducibility constraint, not a formatting preference.
The three provisions that bite a finance pipeline, ✅ verbatim:
- 🚨 "Never derive or infer, or store derived or inferred, information about a X user's: ... Negative financial status or condition", and "Credit or insurance risk analyses" is a listed prohibited use.
- 🔑 The carve-out everything else rests on: "Aggregate analysis of X content that does not store any personal data (for example, user IDs, usernames, and other identifiers) is permitted", subject to the rest of the Agreement. Aggregate, and do not store identifiers.
- 🚨 "X prohibits any use of the X APIs and/or X Content to fine-tune or train a foundation or frontier model with the exception of Grok."
⚠️ The Academic Research product track is absent from the 2026-09-10 docs - searched,
not found. "Retired in 2023" is secondhand and was not verified. v1.1
statuses/user_timeline is still listed in the endpoint map with no deprecation marker,
described as "Limited support; use v2 for new projects."
| item | ✅ read 2026-09-10, support.reddithelp.com |
|---|---|
| Rate limit | "100 queries per minute (QPM) per OAuth client id", averaged "over a time window (currently 10 minutes)" |
| 🚨 Unauthenticated | "Traffic not using OAuth or login credentials will be blocked, and the default rate limit will not apply." Blocked, not throttled. |
| 🚨 Academic research | *"Can I perform research using Reddit developer tools and services? No."* The only sanctioned route is Reddit For Researchers; using the API for academic research "is a violation of our policies." |
| Commercial use | "You cannot use any Reddit developer tools and services for commercial purposes without first getting our permission"; "we'll require a contract"; the listed examples include "Services, research, or data access for fees" |
| Model training | "You may not use content on Reddit as an input for any model training without explicit consent from Reddit." |
| Deletion | "You must remove any user content in your possession that has been deleted from Reddit." 48 hours is a "strongly recommend" sweep interval, not a grace period, and retention "even if disassociated, de-identified or anonymized" violates. |
| Pushshift | reinstated for "verified Reddit moderators" with "use of Pushshift ... limited to moderation use cases only". No public, researcher or developer path exists. |
🔴 redditinc.com/policies/data-api-terms could not be reached from this environment.
Everything above is the help centre - official, but not the contract. If exact Terms
language is load-bearing for you, open it yourself.
🔑 The deletion requirement is the same fact as §4's bias, seen from the other side. You are contractually obliged to destroy exactly the observations whose absence biases your backtest. There is no configuration in which you legally hold a complete sample.
StockTwits
✅ stocktwits.com/developers returns HTTP 200 and renders a page titled "Stocktwits
Subscriptions" - consumer plans (Ad Free $85/yr, Edge $229.50/yr), no endpoint
documentation, no self-serve key registration. The only API route is an Enterprise plan
whose button reads "Get In Touch". ✅ api.stocktwits.com returns HTTP 302 to
stocktwits.com; the historical docs path api.stocktwits.com/developers/docs returns
HTTP 404.
⚠️ hiQ Labs v. LinkedIn - two layers, pointing opposite ways
- CFAA layer: 9th Cir. No. 17-16783, decided 2022-04-18, 31 F.4th 1180. The Supreme Court vacated the first panel opinion and remanded in light of Van Buren v. United States, 141 S. Ct. 1648 (2021); on remand the panel again affirmed the preliminary injunction. 🚨 The panel held hiQ "raised a serious question" whether the CFAA's "without authorization" concept reaches a bot refused access to data that needs no authorization generally. That is a preliminary-injunction standard, not a merits holding that scraping public data is lawful. Writing it as a flat holding - which most summaries do - overstates it.
- Contract layer: N.D. Cal. No. 17-cv-03301-EMC (Judge Edward M. Chen). November 2022 summary judgment: hiQ breached LinkedIn's User Agreement. Consent judgment entered 2022-12-08: $500,000, and a permanent injunction barring hiQ from code for data collection from LinkedIn "using any of the data, source code, or algorithms developed at hiQ", requiring it to delete that code and all LinkedIn member profile data. ⚠️ These terms are quoted from a law-firm client alert, not from the docket - secondhand.
🔑 hiQ won the CFAA question and lost the company on the user agreement. The lesson for a pipeline is that "not a crime" and "not a breach of contract" are different questions, and the second one is the one that ends projects. This library ships no scraper for any of these sources, and will not.
2. ✅ Measured - what reading costs
Arithmetic on X's own published prices, $0.005 per post read:
| budget | posts/year | USD/year | USD/month | % of the monthly cap |
|---|---|---|---|---|
| 1,000/day | 365,000 | $1,825 | $152 | 1.0% |
| 10,000/day | 3,650,000 | $18,250 | $1,521 | 10.0% |
| 100,000/day | 36,500,000 | $182,500 | $15,208 | 100.0% |
| at the 3M/month cap | 36,000,000 | $180,000 | $15,000 | 100% |
🔑 A cashtag-level sentiment panel across a few hundred names is a five-figure annual line item at the low end and hits the Enterprise wall before it becomes a market-wide dataset. Budget it before you design the study.
3. The three timestamps
A post has created_at, the moment your collector could actually see it, and deleted_at. A historical pull shows you the first and never the third, because the post is not there either.
✅ Measured on the seeded panel (INDEX_LAG_PMF sets it): 55% of posts indexed the same
day, mean 0.69 days, p95 2, max 5. And 17.5% of posts are deleted. Names are drawn
by attention rather than uniformly: 39 effective names out of 200, which is why a post
count is not a bet count.
🚨 Nobody in the panel of §4 and §5 has any skill. Every call is a coin flip. Everything those two sections show is manufactured by the sampling alone.
4. 🚨 Measured - deletion survivorship
A post whose call aged badly is deleted more often than one that aged well. A live collector stored it; a historical pull cannot see it. 6 seeds:
| sample | hit rate | Sharpe | sd | % of posts |
|---|---|---|---|---|
| every post (what a live collector had) | 49.9% | -0.103 | 0.440 | 100% |
| surviving posts (what a historical pull returns) | 54.3% | 2.733 | 0.440 | 82.5% |
✅ Measured: a coin flip reads as 54.3% once the losers are gone - +4.4 points from nothing. As a function of how much more often a wrong post is deleted:
| extra deletion on a wrong call | apparent hit rate | true | apparent Sharpe |
|---|---|---|---|
| +0% (the control) | 49.8% | 50.0% | -0.052 |
| +5% | 51.3% | 50.0% | 0.931 |
| +15% | 54.3% | 50.0% | 2.901 |
| +30% | 59.8% | 50.0% | 6.187 |
🔑 At +0% the bias vanishes, which is the control: deletion only matters when it is correlated with the outcome. It always is. And note the size - four points of hit rate backtests at a Sharpe most people would not question. 🚨 That is what volume does. 20,000 posts is enough statistical power to turn a small selection bias into a large, stable, entirely fake edge, and social data arrives in millions. Volume is not a defence against a sampling bias; it is what makes one look like a discovery.
5. 🚨 Measured - the accounts you can name today
Same zero-skill panel. Top decile of accounts by realised call performance, 6 seeds:
| treatment | Sharpe | sd | vs all accounts |
|---|---|---|---|
| all accounts, no selection | -0.058 | 0.608 | - |
| selected out-of-sample (first half, measured on the second) | -0.242 | 0.399 | -0.184 |
| selected in-sample (whole sample, measured on it) | 3.100 | 0.167 | +3.159 |
🔑 Nobody in this panel can predict anything, and in-sample selection manufactures a Sharpe of 3.1. That is what an influencer panel is: the accounts you can name today are the ones that were right, and their track record is the reason you can name them. The out-of-sample split is the only test worth running on such a list, and here it correctly finds nothing.
⚠️ Selecting on the first half is still a trial, and the same argument applies to a list
of managers (../institutional-13f/SKILL.md) or members of Congress
(../congressional-trading-disclosures/SKILL.md). Log it in
../../../fin-core/skills/backtest-validation/scripts/trial_ledger.py.
6. ✅ Measured - the timestamp A/B
This one needs real skill in the panel or there is nothing for the lag to destroy: 25% of accounts, a 6-point edge, a 5-day horizon. 6 seeds:
| key | Sharpe | sd | gap | retained |
|---|---|---|---|---|
| created_at (look-ahead) | 1.270 | 0.484 | +0.257 | - |
| indexed_at (honest) | 1.014 | 0.456 | - | 79.8% |
🔑 The lag is under half a day and it costs a fifth of the signal, because the horizon is
five days. It is the same arithmetic as the 45-day congressional lag against a 21-day
information half-life, at a hundredth of the scale - what matters is lag / horizon, and
a social signal has the shortest horizon in this plugin.
7. Traps
- 🚨 Backtesting on a historical pull and trading live. §4. The two samples are not the same object, and the difference is signed in your favour in the backtest.
- 🚨 Keying on
created_at. §6. Use the timestamp your own collector wrote. - 🚨 Picking the accounts after seeing their record. §5, +3.159 of Sharpe from nothing.
- 🚨 A vendor's "historical social sentiment" file. Ask two questions before anything else: was this collected in real time or pulled later, and what is the deletion policy for records whose source post was removed. If the answer to the first is "pulled later", §4 applies to the whole file.
- 🚨 Bot and coordinated-campaign contamination. A cashtag stream is a target, not an observation. Volume spikes are the most manipulable input in this plugin, and a pump-and-dump generates exactly the pattern a naive sentiment signal buys.
- 🚨 Counting posts as independent. §3: 39 effective names out of 200. Post counts overstate the number of bets by an order of magnitude.
- ⚠️ Storing identifiers. X's aggregate carve-out is conditioned on not storing user IDs or usernames; Reddit requires deletion propagation that anonymisation does not cure. A design that stores a per-account panel is the design that breaks both.
- ⚠️ Assuming a free tier because a tutorial from 2023 used one. §1: there is no free tier in X's current docs, and Reddit blocks unauthenticated traffic outright.
- 🚨 Treating hiQ as permission. §1: it was a preliminary-injunction ruling on the CFAA, and the same plaintiff then lost on contract and paid $500,000 and deleted its algorithms.
8. Scripts and where this sits
scripts/social_feeds.py - the whole access position as ACCESS, with BLOCKED and ABSENT
rows marked; X_PRICES_USD, read_cost() and cost_table(); HIQ; make_panel() with
three timestamps, attention_weights() / effective_names(), index_lag_summary();
survivorship() and survivorship_grid(); panel_selection(); timestamp_ab().
numpy + pandas, seed 20260910, under 2 s, no network, no file writes, no scraper.
../congressional-trading-disclosures/SKILL.md- the same lag argument at 45 days, and the same in-sample selection trap applied to members rather than accounts.../insider-form-4/SKILL.md- the same argument where the timestamp is a regulated filing, so the lag is knowable rather than a property of your collector.../institutional-13f/SKILL.md- selection over managers, and what a partial view of a book does to a reconstruction.../../../fin-core/skills/research-integrity-guards/SKILL.md§1, §2 and §5 - survivorship, availability and trial count; this skill is all three at once in one dataset.../../../fin-llm/skills/llm-finance-agents/SKILL.md- if the plan is to hand posts to a model rather than to a sentiment score, the evidence on whether that survives costs and leakage lives there.../../../fin-core/skills/backtest-validation/SKILL.md- a Sharpe of 2.7 from a coin flip is the whole reason the deflated-Sharpe machinery exists.../../../fin-core/skills/signal-construction/scripts/assert_causal.py- the mechanical check that a feature at time t uses nothing after t.