Prediction moves from benchmarks to real outcomes
A forecasting model that needs no retraining, seven models calling an IPO close, and a specialist that beat the generalists on 318 real website tests.
- Forecasting Company's product handles data preparation and context engineering at inference time, changing the context window is a request choice that requires no new model artifact, and the system requires source versions to be available at each cutoff to prevent data leakage in backtests. m:…5866
- Kilo Code says Anthropic priced its IPO at $83 a share, its first trade printed at $111.50, the stock closed its first day at $97.17, all 14 model runs produced point forecasts higher than the actual closing price, and Gemini 4 Neon and Claude Fable 5.2 were the most accurate, missing the close by 33 cents. Self-reported m:…7059
- In a preprint, Squoosh evaluated UX-1 on 318 historical website A/B tests, UX-1 correctly classified 267 of 318 outcomes with an accuracy of 84.0%, and UX-1 simulated 1,000 shoppers per arm during inference. Self-reported m:…6402
- Gong Labs says Thanksgiving week has 42% fewer sales meetings than a normal week in the US, the week before Christmas closes 102% more new-business deals than a normal week, and deferral language appeared in nearly 50% of calls by the week of December 15. Self-reported m:…9004
Accuracy is not the product
Forecasting Company's product handles data preparation and context engineering at inference time, which is a quiet change with large consequences. Traditional ML forecasting models are trained on large historical datasets but receive only a short recent window at inference, so the model that was trained on years of data is asked to decide on weeks of it. Changing the context window is a request choice that requires no new model artifact, which means the question can change without retraining the model, and users can connect a SQL database or send Parquet files to receive forecasts. m:…5866
The system handles backtesting and output delivery automatically, historical backtests use the same preparation rules as live forecasts, and the system requires source versions to be available at each cutoff to prevent data leakage in backtests. That last rule is the one that separates a serious forecasting tool from a demo: if the backtest can see data the live system could not have seen, the accuracy number is fiction. Forecasting Company's approach parallels the shift from fine-tuning BERT to using LLMs with in-context learning, and agents select and evaluate context for the model while Urs reads and prepares data, including declared aggregations and joins. m:…5866
Every model overshot
Kilo Code ran the kind of test that forecasting tools usually avoid: a single real event with money on the line. Anthropic priced its IPO at $83 a share, Anthropic's first trade printed at $111.50, and Anthropic stock closed its first day at $97.17. The experiment used seven models: Claude Fable 5.2, Gemini 4 Neon, GLM 5.3, Kimi K3.1 Pro, DeepSeek V4.1 Flash 1017, Nemotron 3.5 Super VL, and Qwen 3.8 Max, and the experiment methodology involved taking the median of two runs per model, so each model got two chances and the middle answer counted. m:…7059
All 14 model runs in the experiment produced point forecasts that were higher than the actual closing price, and Gemini 4 Neon and Claude Fable 5.2 were the most accurate models in the experiment, missing the close by 33 cents. Models that stayed close to the pre-market price performed better than those that added optimism, which is a polite way of saying the models were bullish when they should have been boring. The cost of the first round of the experiment was $9.66, and the cost of the second round of the experiment was $1.38, so the whole exercise cost less than a lunch, and still every model got the direction of the error wrong in the same way. m:…7059
A specialist beats the generalists
Squoosh Bench, in a preprint, offers a better way to think about prediction: test against outcomes that already happened. Squoosh evaluated UX-1 on 318 historical website A/B tests, the evaluated A/B tests ran longer than 21 days, UX-1 is fine-tuned on human browsing data, and UX-1 simulated 1,000 shoppers per arm during inference. Instead of asking a model to guess what users might like, the test asks whether it can predict what users actually did, which is a harder and more honest question. m:…6402
UX-1 correctly classified 267 of 318 outcomes, and UX-1 achieved an accuracy of 84.0%. Qwen3 Next 80B was the strongest general-purpose configuration on outcome accuracy, and Qwen3 Next 80B correctly classified 196 of 318 outcomes, so the specialist beat the strongest generalist by a wide margin on the same tests. UX-1 favored the variant in 126 experiments, and of the 126 variants favored by UX-1, 118 won, which means the model was not just accurate in aggregate but right most of the times it took a side. The lesson is the same one the forecasting piece teaches: the model built for the job beats the model that is good at everything. m:…6402
The calendar is a forecast too
Gong Labs turns the same lens on sales, where the calendar itself is the prediction. Thanksgiving week has 42% fewer sales meetings than a normal week in the US, Day and Black Friday each lose 89% of their meetings, and Christmas week has 78% fewer sales meetings. But the week before Christmas closes 102% more new-business deals than a normal week, and the Wednesday before Thanksgiving closes 80% more deals than a normal day, so the quiet weeks are not dead weeks: the meetings that do happen close at a much higher rate. m:…9004
The language in calls tells the same story before the numbers do. Deferral language appeared in 8.4% of calls in mid-September, deferral language appeared in nearly 50% of calls by the week of December 15, and sellers used deferral language in 35.6% of calls by the week of December 15. UK next-year talk reached 65% of calls by the week of December 15, while in the UK, sales meetings drop only 6% during Thanksgiving week. A forecast that ignores the calendar, like a model that ignores the pre-market price, will be optimistic in exactly the wrong week. m:…9004
What is still unresolved
- Whether t0's inference-time approach holds on longer forecast horizons
- Whether model stock forecasts improve with a second round of reasoning
- Whether UX-1's A/B prediction accuracy holds on tests outside its training data
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.