Decentralised News · Research & market intelligence
The hard part of a news strategy may arrive before sentiment scoring: establishing which version of a story was available to the account, when it arrived, and which price remained executable. Official API documentation makes these distinctions visible.
What matters
- Created and updated timestamps describe the provider’s record; they do not record receipt by your system.
- A historical endpoint can return content and ordering that require revision-aware handling before they are used as past trading inputs.
- Data entitlement, processing time and execution price are separate constraints. A faster model does not remove the other two.
Scope of this investigation
This investigation compares 7 primary-source documents. It separates documented rules and features from DN calculations and hypothetical scenarios. Full source dates, definitions and limitations appear below.
Test the assumptions · Inspect the sources · Read the evidence table
The signal
Alpaca’s news reference distinguishes real-time access from an endpoint boundary fifteen minutes behind the present for users without that access. Benzinga documents article revisions and a removed-news mechanism. These are concrete reasons to preserve the version and arrival record, rather than match today’s article text to its earliest timestamp. [1] [4]
DN Alpha Thesis
A sentiment model can classify a sentence correctly while the trading claim remains invalid. If the sentence was added after the event, reached the account later than the timestamp suggests, or was matched with a price that preceded the decision, a persuasive historical chart may describe an opportunity the strategy never had.
This documentary review does not test a sentiment model or claim a predictive edge. It compares the timestamp and update semantics of official news, filing and market-data documentation, then builds a transparent cost bridge. The practical question is whether a proposed signal is eligible for a trade at all; profitability must be assessed only after that boundary has been established.
The headline has a history, not just a date
Alpaca’s streaming schema contains an article identifier, a headline, content, a source and separate created_at and updated_at fields. Benzinga’s own Newsfeed guide similarly distinguishes the article’s first creation from its last update. Those fields establish that a news item can have a history. They do not, on their own, supply every historical text version. [2] [4]
Consider an explicitly hypothetical sequence. At 12:00, a headline reports that a regulator is reviewing an application. At 12:07, the story is revised to say approval has been granted. A dataset downloaded later contains the stronger statement alongside the original creation time. Feeding that text into a simulated 12:00 decision gives the model information it did not yet have. The error survives even if the model and subsequent price calculation are flawless.
The appropriate unit of evidence is therefore an observed version of an item, not merely a URL. Preserve the provider’s article ID, its update timestamp, the received text or an auditable permitted representation, and the local receipt time. If retention rights restrict storage, resolve that constraint before promising a reproducible historical study; this review has not assessed any customer’s licence.
An old article appearing again can mean an edit, syndication, a delayed delivery or a genuinely new event. Those cases should not automatically become four independent bullish signals. A rule for deduplication and revisions is part of the strategy. Changing it after looking at returns changes the experiment.
Access rights are part of the clock
Alpaca’s news endpoint documents different default end times depending on whether the user has real-time access to the feed: the current time with that entitlement, or fifteen minutes earlier without it. The same reference says results are sorted by updated date. This is a documented access and ordering condition, not a measured claim about universal network latency or a statement that every free plan has identical news rights. [1]
That distinction changes what a historical simulation can legitimately assume. A trader using a delayed entitlement cannot justify an immediate historical entry merely because the archive contains a creation timestamp. Conversely, a delayed account is not automatically useless. A strategy operating over hours or days might tolerate that delay if its remaining signal is demonstrated after realistic costs.
For the proposed strategy, write down the actual provider, feed, entitlement and effective date. Then distinguish provider-side availability from receipt by your own process. A subscription description, a REST response and a WebSocket connection are different evidence. A connected socket does not establish that the account is receiving the specific content or entitlement the study assumes.
The commercial decision follows from that separation. Paying for faster data is justified only by evidence that the additional usable time has value after execution costs. A faster feed cannot repair a classifier that mistakes commentary for a new event. Nor can a better classifier recover an entry price that has already moved. The two investments require separate tests.
An update stream must preserve change, not overwrite it
Benzinga recommends delta ingestion for news using updatedSince. Its Newsfeed guide documents creation and update fields and a mechanism for identifying removed items. These are useful controls for keeping a current database current. They do not automatically turn that database into a point-in-time archive of what a strategy previously saw. [3] [4]
For research, keep the current view and the observed history conceptually separate. If a story is corrected, the current view should reflect the correction. The historical decision record should still show which version generated the earlier signal, subject to the provider’s retention and deletion requirements. A removed item may need to be represented by a permitted tombstone or metadata record rather than retained text.
This is particularly important when sentiment is recomputed. Store the model version, its input version, its output and the decision made from it. Re-running a newer model on revised text is a new analysis. It can be valuable, but it does not reconstruct the old agent’s knowledge.
There is also an innocent countercase: many revisions are stylistic or administrative and do not change a signal. A useful audit should measure whether a revision changed the relevant text or decision instead of assuming every update is material. The rule must still be declared before outcomes are compared. Otherwise harmless edits can be ignored when convenient and important ones used retrospectively.
Official-event time can also differ from public availability
SEC documentation provides a helpful example outside commercial news feeds. Its API page says submissions data is updated as filings are disseminated, with a typical processing delay below a second, while XBRL processing is typically below a minute; peaks can take longer. A separate webmaster FAQ describes documents often becoming available on the website one to three minutes after the EDGAR system timestamp and explicitly declines to guarantee that lag. These refer to different stages and should not be treated as competing stopwatch measurements. [6] [7]
For a crypto strategy reacting to a listed company’s filing, the date of the filing, system acceptance, public dissemination, your receipt and the subsequent news summary are distinct events. Joining all of them to the earliest time can manufacture apparent foresight. A government record can be authoritative about its contents while still requiring careful interpretation of availability.
Market data has its own semantics. Binance’s spot trade-stream documentation labels E as event time and T as trade time. Its candlestick payload separately identifies interval start, interval close and whether the candle is closed. None of these fields is your local receipt timestamp. [5]
Do not use the final close of a candle as though it were known at the candle’s opening. Do not infer an executable purchase from a historical trade that occurred before the news arrived. These are logical constraints, not evidence that a particular exchange or data provider is defective. A sound join must preserve what each timestamp actually measures.
A two per cent market move is not a two per cent trade
The calculator uses a long-only, unleveraged hypothetical trade to separate the headline reference from the available entry. The reference price is £100, the assumed executable entry is £101 and the later assumed exit is £102. The observed reference-to-exit move would be 2%. A trader who could only enter at £101 did not capture all of it.
With £1,000 of pre-fee entry notional, the model buys approximately 9.90099 units. A 0.1% entry fee makes the purchase cash outflow £1,001. A 0.1% exit fee leaves approximately £1,008.89 from the sale. After an additional £1 of fixed cost, profit is approximately £6.89 on £1,002 of total cash outlay, or 0.6877%. Every number in this example is synthetic; none describes a provider’s current tariff or an actual trade.
Under those assumptions, the break-even exit is approximately £101.3033. If the exit is £101 rather than £102, the trade loses £3: the price is unchanged from the actual entry and costs still apply. That countercase is why a correct bullish classification is not sufficient evidence of a profitable decision.
The timing inputs are deliberately separate. Ten seconds from publication to receipt, two seconds of processing and one second to the execution decision produce a thirteen-second signal age. The tool does not estimate a price change from those seconds. Users must supply the eligible entry price themselves. Any model that turns delay into predictable lost profit would require additional empirical evidence that this article does not claim.
What would make a stronger sentiment study credible?
The next useful study would define its event universe before looking at returns. It would retain the news versions actually observed, the entitlement in force, gaps in collection, the model configuration and the policy for repeated or corrected stories. It would state whether symbols were assigned at arrival or enriched later. Those choices determine what the model was allowed to know.
A comparison should then hold the decision problem constant. Compare the sentiment rule with a no-trade baseline and a simpler rule using the same arrival constraint, eligible prices, position limits and costs. A more complicated model earns its place only if the comparison demonstrates a relevant improvement; complexity is not evidence by itself.
Separate model selection from final evaluation. A rule tuned on a period and then judged on the same period answers a different question from a rule frozen before a later observation window. If the result disappears after accounting for revisions or realistic entry prices, report that result rather than widening the universe until a favourable chart returns.
This article has not conducted that study. Its contribution is to identify documented constraints that such a study must respect and provide arithmetic that exposes the distance between a headline-price chart and a funded trade. Readers can use it to evaluate a provider’s evidence request without accepting either the claim that news trading is easy or the opposite claim that all delayed information is worthless.
The reader’s decision is about evidence and cost
Before paying for a feed or delegating a trade, ask for one complete example: the exact article version, the account’s first receipt, the resulting signal, the first eligible order and its actual outcome. If the provider only supplies a creation timestamp and a subsequent price chart, the central trading claim remains unresolved.
The existing DN signal-freshness guide helps inspect stale or incomplete inputs. This analysis adds the content-history problem: a fresh delivery can contain an old event, a revision can change the sentence being scored, and a historical download may not preserve the earlier account’s information set. Accurate clocks are necessary, but the text attached to them must also be the right version.
The strongest practical outcome may be a slower, better-defined research workflow rather than a faster trading agent. A user whose decision horizon is long can focus on material revisions and primary documents; a user pursuing seconds must produce much stronger arrival and execution evidence. In either case, treat sentiment as an input to a falsifiable decision rule, with costs and availability established before profitability is claimed.
The evidence, side by side
| Primary source | Documented field or boundary | What it does not establish |
|---|---|---|
| Alpaca news endpoint | Entitlement-dependent default end time; sorted by updated date. [1] | Your historical receipt time or the rights attached to another account. |
| Alpaca news stream | Article ID, created_at and updated_at. [2] | A complete archive of every text revision. |
| Benzinga documentation | updatedSince delta ingestion; created/updated fields and removed items. [3] [4] | That a current database reproduces the past information set. |
| Binance spot streams | Trade, event and candle timing are distinct fields. [5] | An executable order at the historical print. |
| SEC dissemination guidance | API processing and website availability describe different stages. [6] [7] | A guaranteed end-to-end delay for a trader. |
News Signal to Executable Trade Bridge
Synthetic worked trade. No historical signal test or executed trade is claimed.
The interactive calculator starts when this page loads. The worked examples remain available in the article.
Reference price, entry price and entry notional strictly positive; exit price may be zero. Fees are fractions from 0 inclusive to 1 exclusive. Fixed cost and delays are non-negative; all inputs finite.
No model accuracy, fill probability, signal half-life, taxes, borrow or leverage is estimated. An entry price is an assumption unless backed by an actual execution or defensible historical fill model.
How the calculation works
signal_age: receipt_delay + processing_delay + route_delay
units: entry_notional / entry_price
total_outlay: entry_notional × (1 + entry_fee) + fixed_cost
net_pnl: units × exit_price × (1 − exit_fee) − total_outlay
net_return: net_pnl / total_outlay
break_even_exit: total_outlay / (units × (1 − exit_fee))
Scenarios to watch
Illustrative delayed entry
Reference £100, entry £101 and exit £102 produce a 2% reference move but approximately £6.89 net profit, or 0.6877%, on £1,002 cash outlay under the stated fees.
Illustrative unchanged exit
An exit of £101 with the same entry and costs loses £3, despite remaining above the £100 headline reference.
Method and limitations
Read seven official documents from Alpaca, Benzinga, Binance and the SEC on 14 November 2026. This is a comparison of documented data semantics, not a benchmark of providers’ measured latency.
Creation, revision, public availability, receipt and execution are treated as separate concepts. The study did not subscribe to a paid feed, gather an event panel, run a sentiment classifier or perform a held-out backtest.
The worked trade uses synthetic prices and costs. The calculation starts from executable entry assumptions rather than inferring achievable returns from a news timestamp.
Current documentation can change, and an account’s contract may differ from a general endpoint description. Verify actual entitlement and retention rights before collecting or replaying content.
The SEC pages were read at the cutoff but have older publication/update dates. Their timing descriptions are not current measured latency guarantees.
No measured signal accuracy, delay distribution, live fills or return forecasts are supplied. The calculator is a long-only arithmetic illustration.
Primary sources and evidence register
Source dates and reporting periods differ. Several documents from one issuer are one source family, not independent confirmations. The evidence download contains the observations used in this article.
[1] News articles — Alpaca API ReferencePublication date not established; see reporting period · Undated documentation; accessed 14 November 2026
- Default end depends on real-time entitlement; without it defaults to 15 minutes before current time.
- Articles sorted by updated date.
Interpretation limit: Official documentation; documentary review only. No independent operational test was performed.
[2] Real-time News — AlpacaPublication date not established; see reporting period · Undated documentation; accessed 14 November 2026
- Stream schema includes ID, created_at, updated_at, text and source.
Interpretation limit: Official documentation; documentary review only. No independent operational test was performed.
[3] Home — Benzinga API DocumentationPublication date not established; see reporting period · Undated documentation; accessed 14 November 2026
- Recommends delta ingestion through updatedSince for news.
Interpretation limit: Official documentation; documentary review only. No independent operational test was performed.
[4] Mastering the Newsfeed API: Tracking Real-Time Headlines and Price Drivers in Python21 November 2025 · Provider tutorial published 21 November 2025; accessed at cutoff
- Distinguishes first creation and last update.
- Documents a removed-news endpoint and update polling.
Interpretation limit: Older provider tutorial; no current tariff or performance claim is inferred.
[5] WebSocket Streams — Binance Spot API documentationPublication date not established; see reporting period · Current master documentation accessed 14 November 2026
- Trade stream E and T denote event time and trade time.
- Kline payload distinguishes interval boundaries and closed status.
Interpretation limit: Official documentation; documentary review only. No independent operational test was performed.
[6] EDGAR Application Programming Interfaces (APIs)8 April 2025 · Originally 6 June 2024; explicitly last reviewed/updated 8 April 2025
- Updates follow dissemination; typical submissions processing below a second and XBRL below a minute.
- Peak periods may take longer.
Interpretation limit: Official description, not a guaranteed latency or a new 2026 measurement.
[7] Webmaster Frequently Asked Questions — SEC19 May 2021 · Page publication date; accessed 14 November 2026
- Website availability often 1–3 minutes after EDGAR system timestamp, with no guaranteed lag.
Interpretation limit: Older living FAQ; timing scope differs from post-dissemination API processing.
Frequently asked questions
Can created_at be used as my trading entry time?
Only if separate evidence establishes that the relevant version was available to your strategy then and the price was executable. A provider’s creation field does not itself establish either condition.
Are news updates always a source of bias?
No. Some changes are immaterial. Bias arises when information added later is treated as though it was available earlier. Record versions and specify how revisions affect the decision.
Does the article show that paid real-time data is profitable?
No. It shows that entitlement affects the feasible information set. Whether faster data adds enough value to cover its cost requires a separate comparison using actual arrival and execution evidence.
Why does the calculator not turn seconds of delay into a lost return?
That relationship must be estimated from a suitable dataset and may vary across events. The tool keeps time and entry-price assumptions separate so it does not invent such a relationship.
