Help & glossary

Reading a signal

Each tracked asset carries one current call, refreshed twice a day. A call is made of four parts:

• The verdict — ▲ Buy, ▼ Sell or ■ Neutral, with « strong » variants when the score is extreme. The same three marks are used everywhere on the site : on charts, ▲ marks a switch to Buy, ▼ a switch to Sell, ■ a return to Neutral.
• The technical score — a number from −100 (strong selling pressure) to +100 (strong buying pressure), computed server-side from price history alone. Deterministic : the same data always gives the same score.
• The confidence level — 0 to 100 %. Below 50 % the badge is dimmed and dashed on purpose : a shaky call says so instead of dressing itself up as obvious.
• The source — ia when an AI weighed the numbers and issued the verdict, technique when the AI was unavailable and the deterministic score alone took over. The site never hides which one you are reading.

The AI never computes anything. Indicators are calculated in PHP on the server; the model only interprets the resulting numbers — a language model is a poor calculator, and giving it exact figures removes that whole class of error.

To see all of this on a live example, open any asset and follow « Why this call? » : the current verdict is broken down into its exact score components, the model's own words, and the news it could see.

What Gloppr looks at

A single reading of a market is easy to fool. Gloppr combines several independent dimensions, each answering a different question about the same asset, so that agreement between them means something and disagreement is visible rather than averaged away.

The dimensions are chosen to be independent, not numerous. Adding a second measure that says the same thing as the first does not strengthen a reading — it makes a repetition look like a confirmation, which is worse than having neither. Before a new input is admitted, it is measured against those already in place; if it largely repeats one of them, it is left out.

Not all of them carry the same weight of evidence, and the site says which. Trend, momentum, volatility, participation and horizons are computed from market data and are what the deterministic reading rests on. News, cross-asset context and positioning are supplied as context : they have been measured against what prices did next, and none of them showed a usable lead. They are shown because a reader deserves to see what a headline said — not because the site can prove they help.

Anything a dimension cannot supply arrives marked not available, never filled in. A model deprived of a figure does not report the gap — it invents a plausible number, and this site has paid for that once already.

From dimensions to a call

The market-data dimensions are combined into a single score running from −100 to +100, computed by the server on a fixed rule that never changes between assets and never changes with the mood of the market. Some dimensions pull toward mean reversion — a move stretched too far tends to come back — and others toward trend following. The mix is deliberate : on its own, either family is wrong in exactly the situations where the other is right.

That score is then handed to a language model together with the context dimensions. The model does not compute anything : every figure is calculated by the server and given to it. Asking a model to compute an indicator invites arithmetic mistakes that are invisible in fluent prose; giving it the right number removes the whole class of error. Its job is to read the picture, weigh the disagreements, and write the call.

If the model is unavailable or declines, the deterministic reading alone becomes the call, and the signal is marked technique on the page. The site never depends on a model being reachable, and it always says which of the two produced what you are reading.

Where the data comes from

Crypto pairs are read from public exchange data, gathered by scheduled jobs that run on their own clock. Opening a page never triggers a computation, a model call or a network request : what you see was produced earlier and stored. That is why the site is fast, and why a figure carries the time it was taken rather than the time you looked.

US-listed instruments — indices, stocks or ETFs added to the watchlist — come from a market-data provider instead, on the same hourly basis as crypto, with the same dimensions. They read hourly since 1 September 2026 : before that they were scored on daily data, which meant two scores sat side by side on the dashboard while one looked at three weeks and the other at two years. On one stock, the same day, that was the difference between a neutral and a buy.

One thing stays different for them : their call is the deterministic reading only. The model's brief is written for crypto markets, so it is never consulted for a stock or an index, and their asset page says so.

What the site collects and does not use

Several series are gathered every day and deliberately kept out of the analysis. They are listed here because the gap between « collected » and « used » is where most of the honesty of a site like this one lives.

• Series that were tested and did not hold. One of them looked like a credible warning sign for years ; measured against what actually followed, nearly nine alerts in ten led to nothing. It is still collected, and still unused.
• Series with too little history to judge. Some data is only retained for a few weeks by whoever publishes it : an hour not collected is lost for good, so the site gathers it now and will decide later — the reverse order would make the decision impossible.
• Series that are measured and still on the bench. They are scored, confronted with what the market did next, and wired into nothing. On that measurement, none has yet been told apart from a coin toss.

The rule behind all of this is short: collecting is not connecting. A series enters the analysis when it has been shown to add something, and not before — nine ideas have been dropped this way after being measured.

Measurement campaigns

The site runs its own measurement campaigns in the background, separately from what it publishes. They replay decisions on past data with a strict cut-off, compare policies against doing nothing and against simply holding bitcoin, and confront each signal with the asset's own usual drift rather than with zero — because in a falling market any sell call looks right.

Their results are not published here, for one reason: so far none of them has demonstrated an edge. They are how the site finds out what does not work, and most of what has been tested falls in that category. A campaign that produced a flattering number on a handful of dates would prove nothing anyway — one favourable sequence is not a measurement.

Credits

Reading the site costs nothing and requires no account. One action does consume a credit: asking for a portfolio analysis with its reallocation proposal, which is the only click that sends your positions to a language model.

One credit funds one analysis, and a run that fails costs nothing — the credit is taken before the call and returned if nothing comes back. New accounts start with a small allowance; your balance is shown next to your name in the top bar.

How calls are verified

Every call is later confronted with what the market actually did, at fixed horizons of 1, 3 and 7 days. This verification covers crypto pairs only : US-listed instruments carry a technical-only reading and are not part of the track record. Three methodological rules keep the measurement honest :

• Comparison to the asset's own baseline drift. In a market that loses 38 %, any Sell call looks right. So the question is never « did the price fall? » but « did it do better or worse than this asset's usual drift over a window of the same length? » — the median of the rolling windows over its last ~120 daily closes. For a directional call, a gap below 0.5 points is recorded as not significant : noise, not signal. A Neutral call is judged by its own rule — right only when the market stayed within that margin — so its share of misses is not comparable with the directional hit rate, and the track record shows the two families apart.
• Completed candles only. Since August 16, 2026 a call is judged against a finished daily close, never against the still-open candle of the day.
• Criteria frozen in advance. Horizons and thresholds are written once in the code. Adjusting them after seeing the results would be self-validation.
• Correlation is acknowledged — in the arithmetic. Major crypto assets move together : five simultaneous calls are not five independent observations. All directional calls of one day are therefore averaged into a single observation, and the track record reports the resulting daily mean with its z-score (|z| below 2 is noise). It still refuses to conclude anything below thirty distinct days.

Each outcome is written once and never edited — misses included. The running tally is kept internally, misses included.

Glossary

Base 100

A way to draw several assets on one chart : every curve starts at 100 on the first day of the window, so what you compare is percentage drift, not position size or price level.

Baseline drift

How much an asset usually moves over a window of a given length, measured on its own history. The reference every verified call is compared against.

Buy-and-hold BTC

The return you would get by simply holding Bitcoin over the same window — no analysis, no model, no decisions. The honest benchmark for any strategy : beating inaction is not enough, since inaction is just one particular portfolio.

Concentration

The weight of your largest position in the portfolio's total value. A high concentration means the portfolio's fate tracks one asset.

Confidence

0–100 %, attached to every call. From the AI when it issues the verdict; equal to the absolute technical score in the deterministic fallback. Below 50 % the badge is visibly dimmed.

Drift (30 / 90 / 180 days)

The percentage change of an asset over the stated window, part of the long-term read given to the model.

Granularity

The time step of a chart : daily closes for long windows, intraday steps for short ones. Portfolio history switches automatically with the selected period.

Horizon

The fixed delay — 1, 3 or 7 days — after which a call is scored against the market.

Measurement vs reconstruction

A measurement is a portfolio value recorded live and never changed. A reconstruction is recomputed after the fact with current quantities — useful, but it silently becomes wrong if positions change. The site stores which is which and never lets one overwrite the other.

Pegged asset

A stablecoin holding parity with the dollar. Detected by behaviour rather than by a list of names — a list would miss a new one and mislabel an asset merely passing near a dollar. Scored zero, always Neutral.

Retrospective strategy (backtest)

A proposal generated at a past date with price data strictly cut off at that date. Useful to test the machinery — but the model still knows what markets did afterwards, so it proves nothing and is always kept apart from real strategies.

Signal / call

The current verdict on one asset : ▲ Buy, ▼ Sell or ■ Neutral (plus strong variants), with its score, confidence and source.

Source (ia / technique)

Who issued the verdict : the AI reading all the evidence (ia), or the deterministic score alone when the AI is unavailable (technique). Always displayed.

Sparkline

The small 30-day price curve next to each asset — blue when the window ends higher than it started, red otherwise, like every chart here.

Technical score

The −100…+100 number built from bounded indicator contributions (see above). Deterministic and computed server-side.

Tone (news)

Each article is scored from −10 to +10 per asset by a lightweight model before it reaches the analysis — « Bitcoin plunges, Ethereum holds » would be wrong with a single number. It is shown everywhere as a small pill next to the article : −6 reads « clearly negative for this asset », 0 is a genuine neutral, and not assessed means the article was never qualified — an unknown is never displayed as a misleading zero. Promotional pieces are dropped entirely.

« Why » pages

Every asset links to Why this call : the current verdict broken down into its real ingredients — the score's exact components, what the model wrote, the news it could see. Every strategy links to Why this proposal, same idea for a reallocation. Neither shows decision percentages : the model does not expose a weighting, and inventing one would be a lie.

Weighted score

The average of your assets' technical scores, weighted by each position's share of the portfolio value. A marginal line barely moves it; your dominant line mostly decides it.

What this measures — and what it cannot

Almost everything above transforms past prices. What does not — press tone, macro series — has been measured against what followed, and it tracks the move rather than announcing it. None of it predicts : a reversal is observed, then absorbed — never anticipated. Central-bank announcements, regulatory shifts and geopolitical shocks are decided elsewhere, and they reach none of the market data this site reads before they happen. That is a limit of the field, stated here on purpose. The disclaimer at the bottom of every page applies to this one too.