cutebear

a pre-market analysis system that runs before every weekday open. two frontier models read the same evidence without seeing each other, a third merges them with an explicit divergence check, and the next day it grades its own call against what the market did. multi-model systems are easy to demo and hard to keep honest. a market does not care what you claimed.

open cutebear → the method behind it

sign-in required. informational only, and not financial advice.

what it is

a multi-model pre-market read on the nasdaq-100, with self-scoring

runs

9:14am ET weekdays, sixteen minutes before the open

grades itself

6:30pm ET, against the actual closing direction

evidence per run

24 news feeds, an economic calendar, and 31 market instruments

models

two independent frontier models, plus a third synthesis pass

how a morning run works

gather evidence in parallel

24 news feeds are fetched concurrently across four tiers: major financial wires, market-specific desks, macro and central-bank coverage, and specialist sources. an economic calendar is pulled and each release tagged against forecast as a beat, a miss, or in line.

compute the signals before the model sees anything

31 instruments are fetched in parallel and turned into ten derived signals in code, not by the model, among them volatility term structure, tail-hedging skew, leveraged-ETF flow ratios, relative volume, market breadth, index crowding, credit risk appetite, and position against the 200-day average. the model receives verified numbers rather than raw prices to do arithmetic on.

two independent reads

the same evidence pack goes to two models from different vendors, called concurrently, neither able to see the other's answer. one runs with extended thinking enabled. this is the whole point: correlated errors are invisible when you only ask once.

merge, and measure the disagreement

a third pass synthesizes both reads into a single sentiment split, deliberately run on the other vendor's model so the merge does not simply reinforce its own earlier answer. the spread between the two analysts sets the confidence label directly: agreement within five points is high confidence, a gap over ten is low.

degrade honestly

if the merge fails, it falls back to the other vendor's merge; if that fails, to a single model's raw read, and the run records which path it took. a partial answer that says so beats a confident answer that quietly lost half its inputs.

grade the call

every prediction is stored with the trading day it refers to, then graded after the close against actual direction and kept on a permanent record with a running letter grade. predictions that are not yet resolved are marked pending, not quietly dropped.

the running accuracy is not on this page, it is in the system, computed live from graded history and sitting at the top of the log tab. sign in and read it off the same record we do. a hit rate retyped onto a marketing page is a claim; the log is evidence.

what it demonstrates

this is a small system, and that is the point. it shows the practices we bring to much larger client work in a form you can inspect in one sitting.

independence before synthesis

two models that cannot see each other, merged by a third pass on a different vendor. the same principle as scoring competing systems blind in a benchmark.

disagreement as a first-class output

confidence is derived from how far the two reads diverge, rather than asserted by the model about itself. self-reported confidence is close to worthless. measured spread is not.

deterministic work stays in code

every number that can be computed is computed before the prompt is built. models are used for judgment, not arithmetic.

a scoreboard it cannot argue with

predictions are graded automatically against reality on a fixed schedule. the same discipline as holding out an eval set instead of admiring the training loss.

under the hood

php postgres parallel http (curl_multi) google sign-in scheduled cron pipeline 30s quote cache intraday charting

interface

five views: synopsis, full analysis, the news digest it read, the raw signal table, and a graded history log

live ticker

17 symbols with change windows from one hour to ninety days, refreshed on a short cache

access

google sign-in required; every data endpoint refuses unauthenticated requests

want a multi-model system that keeps score?

ensembles, judge harnesses, and evaluation loops for your domain, built the same way.

start a conversation → services