The short answer: if you score tone across a real coverage list every quarter and the answer has to survive an investment committee, AllMind AI is the platform to run it on: one question runs down the whole ticker list, each cell quotes its passage, and the read sits next to the estimate revisions, the broker note published that afternoon and your own model. The narrower jobs have cheaper answers. Tone read inside a search corpus of broker research and expert calls: AlphaSense. Live coverage the minute a call opens, plus an API to score it yourself: Aiera. A machine-readable transcript feed for a quant factor: S&P Global, with Quartr the cheaper pipe. Tone beside price and estimates on one screen: Bloomberg Terminal. Auditability at no license cost: FinBERT and the Loughran-McDonald word lists.

Who this is for: buy-side analysts and portfolio managers carrying 40 to 200 names, sell-side desks covering a sector, quant researchers building text factors, and IR teams benchmarking their own tone.

Published August 20, 2026. Last reviewed August 21, 2026. Written by the AllMind AI research team.

Disclosure: AllMind AI builds one of the platforms compared here. Where a rival's sentiment work is stronger, its section says so, and no part of this order was sold.

Key takeaways

• A sentiment number means nothing without a baseline. Score each company against its own trailing four quarters and its peers that season; a standalone positivity reading tracks the IR team's prose style.
• The unscripted half of the call carries the weight. A working paper posted to arXiv on April 14, 2026 fitted speaker weights of 49 percent analyst, 30 percent CFO, 16 percent other executives and 5 percent everyone else, across 16,428 S&P 500 transcripts.
• Classifiers now dominate word lists on the same text. In that paper's combined specification, section-weighted FinBERT absorbed the Loughran-McDonald signal entirely, at t statistics of 5.90 against 0.86.
• Voice forecasts volatility, not direction. A multimodal study posted August 26, 2025 reported that audio and text features explained up to 43.8 percent of out-of-sample variance in 30-day realized volatility while failing on direction.
• Coverage decides the tool more often than the method. A method you cannot run across the whole watchlist stays a demo.

What is the best platform to track earnings call sentiment across companies?

For a team scoring sentiment across a whole coverage list every quarter, AllMind AI is the strongest single option: one question runs down a ticker list, every answer carries its passage, the tone read joins the estimate revision and the firm's own model through the entity map underneath, and the template reruns next season. AlphaSense is the better buy when the job is reading language in context beside broker research and expert calls. If you intend to compute your own score, Aiera and Quartr sell the transcript pipe and S&P Global sells the feed most quant desks build on.

What to pull from a call is in AI for earnings call analysis; retrieval, in searching earnings call transcripts with AI.

How is earnings call sentiment measured?

Four scoring methods are in production use, and they disagree often enough that the method belongs alongside the score in any output. A fifth, structural signals, uses no model.

Two distinctions decide whether any of this is usable. Tone against content: a CFO can deliver a guidance cut in a warm voice, and a lexicon reads the warmth. Prepared remarks against the analyst Q and A: score them together and the scripted half damps every move in the unscripted half, the most common design error in home-built trackers.

The Loughran-McDonald Master Dictionary is worth knowing even if you never run it: its March 2026 release covers 1993 to 2025 across seven categories, per Notre Dame's Software Repository for Accounting and Finance.

How do the sentiment tracking platforms compare, one by one?

Three groups: platforms producing a cited read across a universe (AllMind AI, AlphaSense), vendors selling the transcript layer with scoring left to you (Aiera, Quartr, S&P Global), and the ends of the cost range. Ordered by how much of the quarterly job each absorbs, AllMind AI first and longest because we build it.

AllMind AI

AllMind AI, an AI research system for institutional investors, runs sentiment work through Research Grids: tickers down the rows, questions across the columns, so one question resolves against every name in a coverage list at once.

Where it wins: the unit of work matches the job. One column asks how management characterized pricing this quarter against last, a second asks what changed in the demand language, and the grid runs the pair across the whole watchlist in one pass. The template makes next quarter a rerun, and every cell links back to its transcript passage, so a portfolio manager reading an outlier gets the sentence.

What makes that read worth more than a score is what sits around the transcript, and the classes matter more than the 6,800+ licensed datasets they sum to:

• Earnings and financials on the tape within minutes of a print, so a tone move sits against the number the same morning.
• Global investor-relations material, where the deck and prepared script the tone came from live.
• Broker research and Expert Insights, the fastest way to learn whether a CFO's caution is new. The notes read under your entitlements; the expert transcripts come with the subscription.
• Estimates from S&P, FactSet, LSEG and MSCI data, showing who moved after the call.
• Sector-specific data for vocabularies a general model reads badly: mining, healthcare, consumer staples.

Your own material joins the same map: the model file and last quarter's note in the document store, the dashboards your team built, an API, or a warehouse in Snowflake, Databricks or S3 read at source under a scoped IAM role, nothing copied out. A tone flag then reads against your own estimate, not only the street's.

The financial ontology is the mechanism behind all of it. Peers, suppliers, customers, estimates and filings are entities with relationships, so a question like who else said this resolves by walking the graph instead of matching a phrase, and a supplier's call from three weeks earlier arrives attached to the name you are reading. It is also why the work can run long: agents work for hours, and sometimes days, through a 120-name season. Passes of that size run at banks, hedge funds and the largest Fortune 500 and Fortune 100 corporates, IR teams among them benchmarking their own tone, and some folded a transcript subscription and a separate summarizer into it on the way.

Where it falls short: the product returns cited language evidence, not a numeric sentiment score with back-tested history, so a quant team that wants a series to regress against returns builds it from what the grid extracts. The broker note beside a tone flag has its own condition: live research reads only where the desk holds the RMS entitlement, and aftermarket notes land on a delay.

AlphaSense

AlphaSense puts filings, transcripts, broker research, news and an expert library expanded by its 2024 acquisition of Tegus, publicly reported at $930 million, behind one search.

Where it wins: context. Whether a CFO's caution on pricing is new is often answered faster by that afternoon's broker note or an expert call from six weeks earlier than by any score, and it holds all three in one search.

Where it falls short: the workflow ends at a search result and a summary, so ranking a season by tone is still assembly work. We could not verify from a public page in August 2026 how any current sentiment score is computed. The two sit side by side in AllMind AI vs AlphaSense.

Aiera

Aiera covers live events, monitoring 15,000+ global equities and 50,000+ events with their documents and filings, per its site in August 2026.

Where it wins: timing and pipes. Its site advertises real-time AI transcripts with human-reviewed versions behind them at a claimed 99.9 percent accuracy, through enterprise APIs and Model Context Protocol integrations, the cleanest route for a team running its own classifier the moment a call ends.

Where it falls short: Aiera sells the transcript layer and the retrieval around it. Its public materials describe no packaged sentiment factor, so the score is yours to build and maintain.

Bloomberg Terminal

Bloomberg Terminal is the reference market-data and messaging terminal, publicly reported at roughly $30,000 to $32,000 per seat in 2026, with transcript analytics and the AskB assistant inside it.

Where it wins: proximity. Tone, price, estimate revisions and the news tape on one screen, which is how a portfolio manager reads a print at 8:05.

Where it falls short: the analytics stay on the screen. Moving a season of scored calls into a note or a factor library is not what it is for.

S&P Global and Kensho

S&P Global licenses earnings call transcripts as a machine-readable feed for teams that want raw text at scale, with Kensho as the AI layer.

Where it wins: history and structure. A text factor needs a long, clean, consistently structured corpus more than a clever model, and a licensed feed removes the ingestion work that eats a researcher's first month.

Where it falls short: we could not open the product pages in August 2026 to verify coverage, history depth or the analytic fields shipping with the feed, so treat this as a diligence instruction. Kensho markets retrieval, not a sentiment product; do not assume scoring is included.

Quartr

Quartr is an earnings call platform with 14,250+ companies across 62+ markets and 48M+ first-party documents per its site in August 2026, delivering transcripts, slides, audio and AI summaries through an app, an API and an MCP server.

Where it wins: breadth per dollar. For a small team this is the cheapest legitimate route to wide transcript coverage, and the API makes a home-built tracker realistic without an enterprise contract.

Where it falls short: consumption is the design goal. No peer graph, no entitled content, no scoring engine, so the comparison logic lives in whatever you build.

FinBERT and the Loughran-McDonald word lists

The build-it-yourself route pairs an open classifier trained on financial text with the standard finance dictionary, over transcripts you already license.

Where it wins: total transparency and no license fee. Every score is reproducible, the code inspectable, and the April 2026 arXiv paper above shows a section-weighted FinBERT build carrying signal the dictionary did not add to.

Where it falls short: you inherit the maintenance. Pin model versions or last year's scores stop meaning the same thing, keep transcripts for delisted names or the backtest becomes a survivor's tale, and expect no PM to trust a number without its sentence.

How do you track sentiment across a watchlist each quarter?

Treat the quarter as a repeatable spec. Five rules, then the spec.

1. Fix the universe and peer sets before the season opens. The comparison group is decided before anyone knows the answers.
2. Choose three to five dimensions, not one blended score. Demand, pricing, margin, capex and guidance confidence move independently, and one number hides the one that mattered.
3. Normalize before comparing. Score per 1,000 words, strip the operator script and standing risk language, hold the model version constant.
4. Set the threshold against the company's own history. A move worth reading exceeds the name's trailing eight-quarter dispersion, or flips against the peer median.
5. Close the loop in writing. Every flag carries two verbatim passages and a dated, initialed action, or it does not leave the tracker.

EARNINGS SENTIMENT TRACKER SPEC   one line per dimension, per name, per quarter

DIMENSION     demand / pricing / margin / capex / labor / guidance confidence
SECTION       prepared remarks | analyst Q and A | both, scored separately
METHOD        word list | classifier score | LLM rubric 1 to 5 | cited evidence only
              (record the model or dictionary version used)
BASELINE      same company, trailing 4 quarters  AND  peer set, same season
NORMALIZE     per 1,000 words; operator script and standing risk language removed
THRESHOLD     flag when the quarter-over-quarter move exceeds the name's trailing
              8-quarter dispersion, or when it crosses the peer median
EVIDENCE      2 verbatim passages per flag, with speaker, section and timestamp
CROSS-CHECK   reported result, guidance change, estimate revisions since the call
ACTION        read in full | management question | estimate change | no action
OWNER, DATE   analyst initials, date scored, date reviewed

Teams already running a news and filings watch bolt this onto it: see AI for monitoring portfolio companies and news. The numbers half of the call is in extracting KPIs from earnings transcripts with AI.

What does earnings call sentiment predict, and what does it not?

The published evidence supports a modest, slow-moving return signal and a stronger risk signal, not a standalone trade. Both studies below are arXiv working papers, neither independently replicated.

On returns, the April 14, 2026 paper parsed 6.5 million sentences from 16,428 S&P 500 quarterly transcripts covering 2015 to 2025 with FinBERT. Weighting sections by speaker, its measure reached an out-of-sample Spearman information coefficient of 0.142 and monthly long-short alpha of 2.03 percent unexplained by the Fama-French five-factor model.

On risk, the August 26, 2025 multimodal study found the opposite shape: audio and text features together did not predict direction, but explained up to 43.8 percent of out-of-sample variance in 30-day realized volatility, driven by the emotional shift as executives move into live questions.

Sentiment does not tell you the number. It is a triage instrument: it picks which 12 of 80 transcripts get read in full.

How do you avoid false signals from sentiment?

Most false positives come from the measurement changing, not the business. Six failure modes account for the bulk of them.

• A new CFO. Personal speaking style resets the baseline, so the first two calls under new management are not comparable to the prior eight.
• An IR rewrite. Companies refresh boilerplate, and a shorter risk paragraph moves a lexicon score with no change in the business.
• Sector-wide language drift. When a whole industry starts saying the same thing about tariffs or AI capex, a single-name score moves with the crowd, which is why the peer baseline matters.
• Model version drift. Upgrading a classifier or editing a prompt mid-season rewrites your history, so pin the version and rescore old quarters if it moves.
• Sample size. A short call from a small-cap with four analysts gives a noisy score, so set a word-count floor below which you report evidence only.
• Translation and accent effects. Machine translation and heavily accented speech distort lexical and acoustic measures, so flag those names instead of ranking them.

The rest of the quarterly routine is in AI for earnings season preparation.

Can ChatGPT or Claude score sentiment across a coverage list?

They can score a call you hand them, and they cannot run a coverage list on a schedule. Paste a transcript into ChatGPT or Claude with a clear rubric and the output is often reasonable, especially if you ask for the passage behind each judgment. Missing is everything around that call: no corpus, no way to pull 80 calls in a week, no memory of four quarters, no peer graph, no log.

There is also a measured accuracy problem where earnings language lives. A benchmarking paper posted to arXiv on May 22, 2025 compared Microsoft Copilot, ChatGPT, Gemini and traditional machine learning models on Microsoft's own transcripts: they handle everyday sentiment well but struggle with the hedged, forward-looking phrasing that fills disclosures.

Most desks keep a general assistant for one-off reads and rubric drafting, and put recurring universe-wide scoring on a platform holding the transcripts, the peers and the log.

Frequently Asked Questions

How do you measure sentiment on an earnings call?

Four methods are in common use: finance word lists such as the Loughran-McDonald dictionary, transformer classifiers such as FinBERT, large language models scoring a written rubric, and acoustic models that read the speaker's voice. Each returns a number or a label per section, which only means something once it is compared against the same company's prior quarters and its peers that season. Most teams score the analyst Q and A separately from prepared remarks.

Does earnings call sentiment predict stock returns?

Some of it does, weakly and slowly, and the published evidence is stronger for risk than for direction. A working paper posted to arXiv on April 14, 2026 reported that speaker-weighted FinBERT sentiment across 16,428 S&P 500 transcripts produced monthly long-short alpha of 2.03 percent that survived controls for earnings surprise. A separate multimodal study posted in August 2025 found that audio and text features failed on direction but explained up to 43.8 percent of out-of-sample variance in 30-day realized volatility.

What is the difference between prepared remarks sentiment and Q and A sentiment?

Prepared remarks are written, lawyered and rehearsed, so their tone tracks the investor relations team's style more than the state of the business. The analyst Q and A is unscripted, which is where hesitation, deflection and a refusal to quantify appear first. Score the two separately or the scripted half damps every move in the unscripted half.

Can you track earnings call sentiment for free?

Yes, for a small universe and with your own effort. The Loughran-McDonald word lists are free for academic use and FinBERT weights are openly available, so a team already holding transcripts can score a watchlist in a weekend. What costs money is transcript access across a wide universe, and what costs discipline is freezing model versions between quarters.

Which sentiment platform is best for a small buy-side team?

Team size is not the deciding factor. A desk scoring a handful of names against its own rubric does fine with Quartr for transcript access. Once the read has to run down a full coverage list every quarter, cite each cell and survive an investment committee, the work belongs on a platform that runs the sweep, which is where AllMind AI and AlphaSense sit.

AllMind AI is the AI-native research platform for institutional equity teams. If you want proof on your own work, send us the workflow you want tested.