cutebear
a pre-market analysis system that runs before every weekday open. thirty-one instruments are reduced to ten derived signals in code, the model reads verified numbers instead of doing arithmetic on raw prices, and the next day it grades its own call against what the market did. predictions are easy to demo and hard to keep honest. a market does not care what you claimed.
informational only, and not financial advice.
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.
one read, on numbers it did not have to derive
the evidence pack goes to a single model, gpt-5.6-luna, with the arithmetic already done. it is asked for judgment, not calculation: which signals dominate, what the market is not pricing in, and the strongest argument against its own conclusion.
tie the confidence label to the signals
each call is labelled low, medium, or high against a fixed rule: high means three or more of the precomputed signals point the same way, low means they contradict each other. the label is anchored to that signal set rather than to how certain the model happens to sound.
degrade honestly
if the model fails or returns something that will not parse, the run publishes nothing and records which step failed. the dashboard then says it has no call for the session rather than showing yesterday's. a missing answer that says so beats a confident answer that quietly lost 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.
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.
the same harness every run
confidence defined against a rule
deterministic work stays in code
a scoreboard it cannot argue with
under the hood
want a system that keeps score on itself?
ensembles, judge harnesses, and evaluation loops for your domain, built the same way.