The 5 AI indicators, explained

How fragile is the market right now?

A liquidation cascade is a chain: one forced close pushes price into the next stop, which forces the next close. Whether that chain can start depends on two things: how much leveraged position sits close to the current price, and how much volume the market has been absorbing. Cascade Radar measures both.

Limits. The ladder is a model of where positions would be liquidated, not a feed of resting orders. It says how fragile the market is, not when the push comes.

Where is price being pulled next?

Price tends to visit big pools of liquidations before it reverses: the pool is where forced orders are waiting, and forced orders are what moves a market that nobody else wants to move. The Magnet finds the dominant pool on each side and says which one is pulling harder.

Limits. A magnet is a probability, not a schedule. Big news moves price through both pools without stopping at either.

Which way does the evidence lean, and what is driving it?

Most "AI" indicators are one formula with a name. Market Brain is eight readings and a learning step: it measures which of the eight actually predicted the next six bars on this coin, on this timeframe, and weights them accordingly. A factor that does not work here weights itself to zero.

Limits. Learning from 280 bars means the weights lag a regime change by a few dozen bars. Funding and crowd factors need the exchange to publish them for the pair; where it does not, those two factors read zero and the model says so by not naming them.

When it looked like this before, what happened next?

Market Memory does not model the market. It looks it up. Using the same eight factors as Market Brain, it finds the past bars whose whole state most resembles right now and reports what price did after them. It is the market's own memory: a base rate, honestly labelled.

Limits. The memory is only as long as the loaded history (a few hundred to a few thousand bars). On a coin that listed last month it has little to remember, and it needs 160 bars before it draws anything.

Where has leverage actually been force-closed in the last 24 hours - and what happens when price goes back there?

This one is not a model and cannot be. Every mark is a liquidation that already happened: our own collector records them across eight exchanges as the venues publish them, and the indicator puts each big one on the chart at the price it printed at and on the candle it printed on. Nothing here is projected forward. It is the record of where leverage died, read off the same feed that drives the liquidation heatmap and the Rekt page.

Limits. It covers the ten coins our feed records - BTC, ETH, SOL, XRP, DOGE, BNB, ADA, LINK, AVAX and LTC. On any other market the indicator cannot be switched on at all and says it has no data, rather than drawing an empty chart you might read as calm. The window is the last 24 hours and nothing older is kept on the chart, so a level that mattered last week is not marked. And an exchange only publishes what it chooses to publish: the marks are a floor on what really liquidated, never a total.

Why did price just jump or drop - and which of this month's moves had a dated reason behind them?

Every abnormal move of the last 30 days gets a flag on the candle it started on, and the flag carries the most likely cause in one word. The causes come from MarginPad's Catalyst Engine, which reads exchange announcements (Bybit, OKX, Binance, Upbit, Hyperliquid), SEC filings, the macro calendar, the fast news desks and the official channels with each item's own timestamp - and from our own data: the liquidation feed, the whale tracker, the tape's market orders over a million dollars, the S&P 500 and Bitcoin itself. A move is judged against this coin's own 30-day volatility, never a fixed percent, so 0.6% can be a move on BTC and nothing on a meme coin. Realism is measured, not assumed: for every kind of event we measured how far price moved in the 15 minutes after it against the same market's ordinary bar and a week-earlier placebo, and a kind that has measurably moved nothing (most media headlines) cannot claim a move by timing alone. A language model reads every item for importance, the coins it is really about and the direction it implies. Every flag leaves with a sentence, the unknown ones too: what was checked, what the clock and the market were doing at that minute.

Limits. News must be dated inside 90 minutes before the move (or 30 minutes after, in which case the card says the market led the news); a source that publishes thirty items a day cannot claim a move by time alone, so a transfer alert or a generic headline rarely counts. The cascade test needs our liquidation feed, which covers ten coins; elsewhere only news, whales and BTC can be named. The whale tracker keeps 14 days. Causes are attributed by rules, not read by a model, so a headline that merely mentions a coin can be named for a move it did not cause - the score and the confidence on the card say how strong the match is. Checked against a placebo: shifting every news item by a week collapses the explained share, which is the proof the matching is not chance.