Methodology

How the forecasts and nowcasts on FinSentS are made, how they are checked, and what they cannot do. Every figure here is a statistical estimate, not investment advice.

Who produces them

FinSentS, the Financial News and Sentiment Screener, is published by InfoTrie. The forecasts are computed by InfoTrie's API from the data below, by the rules on this page. The same forecast is shown to every user: nothing uses your holdings, your watchlist or your profile, and a plan only shortens the horizon you see.

Data

• News sentiment and buzz: InfoTrie's daily scores per company and index, from the news it collects, on a −1…+1 scale. The last 180 days feed a forecast.
• Prices: adjusted end-of-day closes of the company's US listing, the last two years, for the volatility forecast.
• Macro releases: the economic calendar, which keeps the first value published for each period. Where revisions are loaded, a nowcast may use the values known on each day instead.

Sentiment and buzz forecasts

Five simple models compete: the last value (a random walk), the window mean, an autoregressive model, a local-level Kalman filter, and the autoregressive model around a day-of-week profile, for series whose weekends differ from their weekdays. A sixth candidate mixes the two models that did best so far; it is served only when it beats the best single model by at least 5 %. Each answers with a distribution, shown as five quantiles: the median, the 50 % and the 90 % intervals.

They are compared on a rolling-origin backtest: at each of the last 30 days, every model is refitted on the data up to that day only and scored on what came next, with the pinball loss of its quantiles. The lowest loss wins. The random walk is the baseline, so the forecast served is never worse than "no change" on the backtest. The page shows the backtest next to the forecast.

Volatility forecasts

The volatility forecast says how large daily price moves are likely to be, not in which direction. A GARCH(1,1) model and its asymmetric variant GJR-GARCH(1,1), where a fall raises the next day's volatility more than a rise, are compared with the historical variance on the last 60 origins, with the QLIKE loss. GARCH is served only when it beats the historical variance; GJR only when it beats both, GARCH by a margin. A scenario replays a hypothetical move of the next trading day through the same fitted model.

Macro nowcasts

A nowcast estimates a figure not yet published (a quarter's GDP, a month's inflation or jobs) from the monthly indicators already out. Several models compete, from the simplest to the most complex: bridge equations, a one-factor model, dynamic factor models (estimated in two steps or by EM, with one or two factors, and a mixed-frequency one that uses the target too), and a combination of them weighted by their past accuracy.

• Point in time. At each date, the models see only what was published by that date.
• Backtest. Each past period is nowcast on the day before its release and scored against the first value published.
• Baseline. An autoregressive model of the figure itself. A model is served only when it beats it.
• Parsimony. A more complex model replaces a simpler one only when its loss is at least 5 % lower, so that noise does not pick the model.
• What moved it. Each day's move is split by indicator: the model served today is re-run on yesterday's data, then with each new release added in turn. The split depends on that order.
• Surprises. The gap between a release and the nowcast of the day before, in standard deviations, is the surprise. Each country's surprise index is their recent average.

Calibrated intervals

A model's 50 % and 90 % intervals are only as good as its assumptions. Each forecast and nowcast is therefore checked against its own backtest errors: if its intervals were too narrow or too wide, they are widened or narrowed by the factor that would have made them hold, learnt at each step from the errors already known then. The recalibrated intervals are served only when that replay scores no worse than the model's own, and the page says which it shows and how often the 90 % interval held.

Track record

Each night (01:30 New York time for companies, 02:30 for nowcasts), the served forecasts are stored. A stored run is never rewritten, and each one is later scored against what happened. The track record is public, on each company and market page, on each nowcast's history page, and for every family at once on the track record page, with the PIT histogram: where the outcomes fell among the five quantiles, against where calibrated intervals would put them.

Technical and fundamental scores

Where shown, a company page ranks the company on seven factors: trend, momentum, low volatility, value, quality, growth and news sentiment. Each factor is a percentile from 0 to 100 against comparable companies in the same sector (the whole universe when the sector is too small), computed each night from what was known that day: prices up to that close, financial statements counted only 90 days after their period ends, and news sentiment as it was published, never revised afterwards.

An overall score is shown only when a rolling-origin backtest finds that a weighting of the factors ranked companies better than an equal blend on the following month's return; news sentiment is included only if it adds to that. Otherwise the factors are shown alone, and the page says so. The technical signals listed (moving-average crosses, RSI levels, breakouts, volume spikes) are dated facts about the price, not forecasts. None of it is investment advice.

AI briefs

Where the AI is on, the nowcasts page carries a short brief per country, written by an AI model from the figures on that page only. A brief that cites a number not among those figures is not shown. It repeats the estimates; it adds no forecast of its own.

Backtests

On the backtest page you define a rule and FinSentS replays it on past data, one trading session at a time. The signal of a session is computed only from what was known before that session opened: the news sentiment of the days before (as first published, unless you choose the values as served today), the closes up to the previous session, and the sentiment forecast stored the night before. The portfolio trades at the close, pays the cost you set on what it trades, and holds until the next close. The universe is the index's members today, applied to the whole period, so companies that left the index are missing: results are flattered by that. The full rules are public (docs/BACKTEST.md in our engineering documentation).

A backtest shows how a rule you chose would have behaved. It is not a forecast, FinSentS does not publish or recommend any strategy, and past results do not predict future ones.

Limits

• Statistical models extrapolate the past. A break the data has not seen will not be forecast.
• An interval is not a guarantee: a 90 % interval is expected to miss about one time in ten, and more often when conditions change.
• The calendar's labels and first releases come from a data vendor. An error there flows into the nowcast until it is corrected.
• FinSentS does not forecast prices or returns.

A statistical estimate, not investment advice.