Source check
Named outlets and original documents, not reposts of reposts. A wire story that cannot be traced to its origin gets no score at all.
Free editorial · AI in investment
Echo Investment Journal
An independent desk reading one narrow corner of artificial intelligence in investment: how machines score the sentiment of financial news.
Free articles, methodical checks, and the awkward cases where mood-reading models misfire. Nothing on this site is for sale — no plans, no signals, no paid tiers.
The subject
What the journal keeps score of
Sentiment analysis converts financial news — headlines, filings, wires, transcripts — into measurable tone. That number then flows into research desks, index notes and strategy papers. This journal is a plain-language record of how the conversion is done, and how often it deserves suspicion.
- Word-list scoring. How lists in the Loughran–McDonald tradition weigh words like risk or withdrawn inside a filing, and why ordinary vocabulary makes financial text a trap for general dictionaries.
- Model scoring. How FinBERT-family classifiers read a whole sentence and return a polarity instead of a word count — and what that buys over the older method.
- Named-entity steps. How a pipeline decides which company a headline is actually about before any score can be attached to a ticker.
- Aggregation. How thousands of single-article scores become one daily sentiment index that a researcher can chart at all.
Each thread is followed the same way: where the data came from, who measured it, and whether the claim stands once taken out of its press release.
The method
Five checks before a sentiment claim runs here
Every article that quotes a machine-read mood passes the same gate before it reaches the page. If a check fails, the claim is cut or marked exploratory rather than established.
Model check
The tool behind every number is named — which lexicon or model, which version, run how. A score without provenance is treated as a rumour.
Baseline check
Polarity in isolation says little. Each value is placed against a neutral reference — an outlet's typical tone, or a moving window — before words like sour get printed.
Granularity check
Headline heat and body text are read separately, because a frightened headline sitting on a calm article is a different creature than it first appears.
Failure check
Sarcasm, recycled alerts and boilerplate warnings are hunted by hand before a number is trusted. Machine readers stumble on exactly these.
The whole gate, plus the threads the desk currently reads, is laid out in the coverage guide.
Failure modes
Where machine reading breaks
Confidence is a model setting, not an insight. Loughran and McDonald's 2011 study showed why general-purpose negative word lists misfire on financial text: words such as liabilities, debt and taxes are routine record-keeping, not omens. A reader that does not know this grades a bond prospectus as a tragedy.
- Irony. A line like great, another downgrade reads positive to many scorers while meaning its opposite.
- Recycling. A republished crash report re-delivers yesterday's shock as this morning's signal.
- Boilerplate. Filings repeat legally required warnings that nobody wrote for tone, and a tone model reads them anyway.
- Overfit backtests. A mood signal that only shines after re-tuning is a story with a good narrator, not a strategy.
The journal documents these breaks as carefully as the industry advertises its successes — a reader who knows where a tool snaps can judge its promises precisely.
The readers
Who the journal is written for — and who it is not
The journal is editorial, free, and sells nothing; its readers are readers, not customers. Two short lists keep that expectation honest.
For hands-on analysts
People who build or evaluate systematic strategies and want to understand what a news-sentiment score actually does before trusting one in a backtest.
For students of computational finance
Readers learning the field who want the pipeline explained in working English before opening the formal papers.
For sceptical news desks
Editors and writers curious how machine scoring reshapes financial journalism — and what to double-check before quoting a sentiment index.
Not for signal hunters
No scores, alerts, portfolios or trade ideas are published here. Visitors looking for a tip line will not find one, by design.
Not for asset gatherers
The desk takes no mandates, holds no funds, opens no accounts. The journal writes about AI in investment; it does not practise investment or give advice.
The invitation
Read along, then interrogate the desk
A journal about machine-read sentiment should accept the interrogation it gives its sources. Questions reach the same editorial desk as topic proposals and challenges to published claims.
Use the inquiry form, the desk line, or the mailbox — whichever suits. The answer may be a correction, and on this subject that counts as a good outcome rather than a defeat.
The editorial desk
Direct line and mailbox — a human on the journal's side reads both.
- Phone
- +886 2 2727 5418
- Desk
- No. 109, Songgao Rd., Xinyi Dist., Taipei City
110 Taipei City, Taiwan
Inquiries only — the journal sells nothing, so there is nothing to buy here.