A weather forecaster doesn't tell you whether it will rain: they tell you how much each force in play weighs, how coherent the picture is and what would change it. Sinottica works the same way on markets.
families of sources
weight per factor
threshold below which nothing goes out
bulletin, Italian time
On-chain and exchange flows, derivatives (funding, open interest, skew), order-book microstructure, analyst revisions, positioning, macro calendar. Frequencies from one minute to one day.
Every series becomes a deviation from its own regime, not an absolute value: funding at −0.012% on crypto, or a −2% earnings revision on a US stock, weigh differently in a calm market and in a stressed one. Dirty data is discarded, not interpolated.
Each factor is given a weight from 0 to 40 based on how it has held up historically on that market and in that regime. Weights are summed by direction: whatever stays below the threshold is, by definition, noise.
The text is written only from the computed weights: no invented factors. The invalidation condition stays monitored until the close, then the read goes into the record.
No family carries a fixed weight: what counts is how it has held up historically on that market and in that regime. On BTC funding often weighs twice what it does on a stock; on NVDA it is order revisions that matter.
“High, medium, low” is the maximum weight that family can reach on the market, not the weight it carries today.
The heavy lifting artificial intelligence does at Sinottica is statistical: identifying the regime, estimating the weights, measuring uncertainty. The language model only steps in at the end, to put an already computed result into plain English — and it cannot write anything the numbers do not say.
A factor enters the model only if it holds up across several regimes and several similar markets. The ones that work on the recent past alone are dropped, however good they look in backtest.
The language model has no access to the markets: it only receives the table of weights already computed. It cannot invent a factor, because it knows none beyond the ones we hand it.
Models are retrained on a regular schedule and compared with the previous version over the same period. If the new one does not do better, it stays on the bench.
Three ingredients: how much the factors agree with one another, how clean the data feeding them is, and how well the dominant factor has held up historically. The result is a number from 0 to 100.
The invalidation condition is a measurable fact, not an impression: a level, a threshold, a number of sessions. When it triggers, the read closes — and it does not count as a miss, it counts as a voided read.
“Voided if funding moves back above zero for 48 consecutive hours, or if open interest on perpetuals climbs back above the levels of 18.07.”
Indices and US stocks example“Voided if the index closes two sessions above 5,480, or if the twelve-month earnings revision turns positive again before options expiry.”
You get the alert the moment the condition is met, not the day after.
Position size, leverage and stops depend on your capital and your tolerance. We do not know them and we do not guess them.
A hack, a sudden bankruptcy or an unexpected political decision are not in the previous day's data. When they happen, open reads are voided.
The minimum horizon is a few days. Below that scale noise dominates the weights and the method stops making sense.
The public record shows how it has gone so far. It is not a guarantee about the future, and anyone who gives you one is selling you something else.
The Monday bulletin is free and contains one full read, weights included.