Forecite[Use cases]Jun '26
A backtest is a claim about what you knew at 07:02
Point-in-time correctness, a score you can take apart, and two horizons that are allowed to disagree. What a systematic desk checks before a news signal is allowed near a live book.
Yin Nguyen Quant Research
[Use cases]May '26
Most mornings the screen should be boring
A PM I talk to runs a $180m book and has until 8:15 to decide what matters. Here's what she actually looks at, what she ignores, and the three places Forecite still can't help her.
Emma Barret Product
[Use cases]May '26
Your agent can't skim
A person reads a wire release and knows in two seconds whether to care. A program needs that judgement as an integer, on a scale that doesn't move, next to a link it can cite. Here's the surface we built for the caller that isn't a person.
Martin Falzer Infrastructure Engineering
Latest articles
A briefing bot built on a SaturdayAn analyst wired his own morning briefing out of the history API, a cron line and about ninety lines of Python. The interesting part isn't the API call. It's the constant on line four.Jun '26[Product]A backtest is a claim about what you knew at 07:02Point-in-time correctness, a score you can take apart, and two horizons that are allowed to disagree. What a systematic desk checks before a news signal is allowed near a live book.Jun '26[Use cases]Most mornings the screen should be boringA PM I talk to runs a $180m book and has until 8:15 to decide what matters. Here's what she actually looks at, what she ignores, and the three places Forecite still can't help her.May '26[Use cases]Your agent can't skimA person reads a wire release and knows in two seconds whether to care. A program needs that judgement as an integer, on a scale that doesn't move, next to a link it can cite. Here's the surface we built for the caller that isn't a person.May '26[Use cases]Lookback is a plan boundaryA Starter key asks for a week of history and is served the 48 hours its plan reaches. Here are the windows, the page-size caps, and how to walk a cursor without guessing where the end is.May '26[Product]Reading the tape in a window we don't ownAn analyst was copying headlines out of our feed and pasting them into a chat window eleven times a morning. Eight MCP tools, one environment variable, and the clipboard step goes away. Three plan limits do not.Apr '26[Product]Most of a model is the labelTwo names can print the same 1.8% hour and only one of them moved. Deciding what counts as price impact is harder than fitting anything to it, and it is the reason our scoring model gets rebuilt on a weekly clock instead of a quarterly one.Mar '26[Inside Forecite]Verify the bytes, not the JSONEvery Forecite delivery carries an HMAC-SHA256 of the raw body in X-Forecite-Signature. Here's a working verifier in Python, TypeScript and Go, and the one framework line that quietly breaks all three.Mar '26[Guides]Warren works on what you hand himAbout a third of what an analyst reads in a week arrives as a forward, and none of it has a badge on it. Warren runs the same Verdict Engine over the thing in your inbox, with one flag set differently and a plan gate people keep tripping over.Feb '26[Product]Every latency number needs a measurement pointA release crosses at 13:30 and the row reaches your socket some number of seconds later. Which number you get depends entirely on which two events you measure between, and most published latency figures quietly measure the shortest span available.Jan '26[Perspectives]A filter you can paste into SlackEvery chip you click on the feed writes itself into the address bar, which makes a view something you can send someone. Here's the parameter model behind it, the place it bites, and the same filter expressed as a curl.Dec '25[Guides]Forecite raises $20M Series AWe've raised a $20 million Series A to accelerate our mission of building the intelligence layer for the markets. Here is what the money buys, who it's for, and the standard we intend to be held to.Nov '25[Company]Sentiment is not a directionA release can read glowing and still be worth nothing. Novelty carries 0.30 of the actionability score, the gate sits at 60, and that ordering is the only reason a tone reading is any use on a wire feed.Oct '25[Perspectives]An empty terminal is the feed workingA key, one socket, and about twelve lines in Python or TypeScript. Plus the four places people stall on the way, including the one where nothing is wrong and nothing arrives.Sep '25[Guides]Four pixels on, six pixels offOur dashed guide lines aren't borders. They're SVG, because CSS can't round the end of a dash, and that decision cost a week we didn't have. Here is what the rest of the drawing is made of, and the three places it stops being honest.Sep '25[Inside Forecite]