project · live · artificial intelligence
sam
Our own forecasting models: sam v1 gives calibrated probabilities, sam v2 forecasts the path of a price, and sam v3 combines them with a candle model. Every forecast is scored in public.
What it is
sam is the family of models qourat builds, trains and runs itself. It answers typed questions (will the next close be higher, where will the price be in three hours, what will the next candles look like) with numbers and probabilities, never with prose, and it runs on our own server. It is trained only on outcomes we own: every forecast on marketprice is settled against the close that really printed, and every claim in a research note is settled by its independent review.
sam v1 is a set of small decision models, one per asset class and horizon. Each reads the recent state of a market and gives the probability that the price will close higher, calibrated on data it was not trained on, and records its accuracy, Brier score and calibration error so nobody has to take the probabilities on trust.
sam v2 forecasts the whole path of a price: a median and a band from the first to the ninth decile for every step ahead, five minutes at a time for three hours and one minute at a time for an hour. It is built on open weights, credited on the marketprice disclaimer page, and trained further on our own bars.
sam v3, live since 2026-09-25, is one system made of four parts: sam v1, sam v2, a candle model that samples the likely open, high, low, close and volume of the next candles, and a combiner that weights each part by its own record on the live scoreboard. The candle model was measured on our own bars before it went in and did not beat a coin flip on direction, and its bands were worse calibrated than sam v2, so it counts for shape only: it draws the wicks and the volume, and it carries no weight on direction or bands until its own record beats sam v2 over at least 100 settled paths. The terminal draws the sam v3 forecast on its chart and labels it a forecast, not a price.
Every part is scored on its own and the blend is scored too. Four live cards, one each for sam v1, sam v2, the candle model and sam v3, sit on the forecast and scoreboard screens of the terminal for every product and frequency. A horizon with fewer than 100 settled forecasts is labelled warming up rather than scored, and a forecast whose outcome never came is void, not a miss.
sam v4 is research, not a product. It would be a layer on top of sam v3 that turns forecasts into decisions: buy, sell or hold, with a position size, a stop and a risk budget. It is built only if a research note shows, from cited sources and from its own tests on our bars, that such a layer beats buy-and-hold and beats sam v3 used on its own, after costs, on data it was not trained on. If the evidence is not there, it is not built and the note says why.
Demonstration live numbers, rebuilt with every publish · 2026-09-25 11:27 UTC
full scoreboard →The oldest of the sam questions, answered by sam v1: every minute, for every open market on marketprice, will the next one-minute close be higher? Below: its score per asset class against the baselines it has to beat, the last twelve settled calls, and what it has been trained on. The sam v2, candle model and sam v3 cards for the longer paths are on the full scoreboard. Nothing here is advice; a minute of price is mostly noise and the numbers say so.
| asset class | settled minutes | sam v1 · accuracy | sam v1 · brier | persistence · brier | always up · brier | coin flip · brier |
|---|---|---|---|---|---|---|
| crypto | 2,287 | 49.5% | 0.2527 | 0.4862 | 0.4998 | 0.2500 |
| metals | 2,756 | 41.5% | 0.2561 | 0.4209 | 0.6299 | 0.2500 |
| forex | 5,504 | 50.6% | 0.2497 | 0.5351 | 0.5274 | 0.2500 |
| stocks | 3,786 | 50.4% | 0.2513 | 0.4923 | 0.4950 | 0.2500 |
The last twelve settled calls
| market | minute | close | sam v1 said | next close | right? |
|---|---|---|---|---|---|
| EURUSD | 2026-09-25 11:25 UTC | 1.1405 | down 54% | 1.14061 | ✗ |
| GBPUSD | 2026-09-25 11:25 UTC | 1.32522 | down 55% | 1.32535 | ✗ |
| PAXGUSD | 2026-09-25 11:25 UTC | 4303.18 | up 63% | 4301.66 | ✗ |
| BTC-USD | 2026-09-25 11:24 UTC | 84941.7405 | down 51% | 84893.785 | ✓ |
| ETH-USD | 2026-09-25 11:24 UTC | 2725.595 | down 51% | 2723.95 | ✓ |
| SOL-USD | 2026-09-25 11:24 UTC | 120.5335 | down 51% | 120.566 | ✗ |
| EURUSD | 2026-09-25 11:22 UTC | 1.14059 | down 54% | 1.14065 | ✗ |
| GBPUSD | 2026-09-25 11:22 UTC | 1.32564 | down 57% | 1.32551 | ✓ |
| PAXGUSD | 2026-09-25 11:22 UTC | 4303.18 | up 63% | 4303.18 | ✗ |
| EURUSD | 2026-09-25 11:21 UTC | 1.14061 | down 54% | 1.14059 | ✓ |
| GBPUSD | 2026-09-25 11:21 UTC | 1.32561 | down 58% | 1.32564 | ✗ |
| PAXGUSD | 2026-09-25 11:21 UTC | 4303.18 | up 63% | 4303.18 | ✗ |
What sam v1 has learned from
- crypto: trained on 4,321 settled minutes; the text reader won the held-out test (accuracy 51.3%, Brier 0.2496, calibration error 0.003).
- metals: trained on 2,488 settled minutes; the text reader won the held-out test (accuracy 78.8%, Brier 0.2237, calibration error 0.253).
- forex: trained on 4,970 settled minutes; the text reader won the held-out test (accuracy 54.0%, Brier 0.2493, calibration error 0.020).
- stocks: trained on 3,777 settled minutes; the text reader won the held-out test (accuracy 49.0%, Brier 0.2503, calibration error 0.023).
- research labels: 804 claim-review verdicts written by the research desk for the question “does this passage support this claim?”. sam takes that question over from the reviewer once 300 exist.
The research behind it 6
Non-Autoregressive Typed Decisions: How Does Laya Compare to BERT for On-Device Classification?
The verified claims do not include any experiment that runs Laya (LAYA, the Layer-wise Attention Aggregator) against a fine-tuned BERT-base model on the same on-device binary or multi-class text classification task with matched latency and accuracy measurements, so no direct answer to the comparison question exists in the evidence. What the claims do show, separately, is that LAYA is a lightweight output head evaluated on image classification datasets, where it reaches competitive predictive performance and gives depth-aware explanations [12], while fine-tuned BERT and its family (RoBERTa, DistilBERT) are documented as strong, low-latency, low-cost classifiers on text benchmarks compared against LLM prompting, not against LAYA [1]. Because LAYA's own reported comparison is against other output heads on image tasks, and BERT's reported comparisons are against LLM prompting and against distilled or quantum-hybrid variants of itself, a builder cannot infer from these claims which one would win on an on-device text classification benchmark. Anyone who needs that answer must run the head-to-head test themselves; this note lays out how, using the mechanisms, numbers, and costs that the sources do report for each side separately.
Kelly Criterion from Calibrated Forecasts: How to Size Positions When p_up Estimates Are Noisy?
No claim in this evidence base gives a single formula that adjusts the Kelly fraction for estimation error in p_up while also pricing in transaction costs; the evidence instead covers these pieces separately: the classical Kelly optimum under fair odds and no costs [3], the fragility of Kelly and high-fractional Kelly to short-term ruin [4], the practice of using a fraction of the Kelly fraction to reduce risk [5], a feedback-control approximation that turns the log-growth problem into quadratic programming [13], evidence that estimation error in inputs (not just p_up) can badly distort optimal weights [8], and separate work on transaction-cost-aware investment and ruin-probability problems [7][10][1]. A builder must therefore combine these pieces rather than cite one paper for the whole pipeline. The honest position is: fractional Kelly, drawdown floors, and cost-aware thresholds are each independently supported adjustments, but no cited study measures their joint effect on ruin probability when p_up itself is noisy and drifting.
Quantum Time-Series Forecasting: Can a Hybrid Model Beat LSTM on a Real Dataset in Simulation?
No claim in this evidence set reports a head-to-head test of a trained quantum-classical hybrid model against an LSTM on the same real-world forecasting dataset, so the question cannot be answered directly from what is here. The closest real-dataset evidence for forecasting shows a classical LSTM strongly beating ARIMA and SARIMA on ten years of renewable-energy data [4]. A separate paper reports a hybrid quantum-classical model underperforming in a fully connected traffic forecasting setting, but the exact figure attached to that finding is truncated in the source material and cannot be quoted as a usable number, so it should be read only as a qualitative result of underperformance [1]. The only clean, complete hybrid-vs-classical comparison available is not forecasting at all: a small variational quantum circuit beats a BiLSTM and logistic regression on a financial sentiment classification task, while a fine-tuned FinBERT beats all three [7]. A builder who wants to answer the original question has to run the experiment; today's literature does not contain a valid direct result, and this note sets out exactly what pieces exist and how to assemble them.
Conformal Prediction for Time-Series Forecasts: How to Wrap Foundation Model Quantiles?
No claim in this evidence set reports coverage or interval-width numbers for wrapping a time-series foundation model's quantiles on financial OHLCV data at 5-minute to daily horizons, and none compares such numbers against raw model quantiles [1]. The mechanisms exist and are well specified: adaptive conformal inference (ACI), conformal PID control (quantile tracking and error integration), weighted conformal prediction, Bellman Conformal Inference, and ensemble methods (EnCQR, EnbPI) all give recipes to turn quantile or point forecasts into calibrated intervals under distribution shift, with coverage guarantees stated as long-run averages rather than pointwise guarantees [1] [2] [4] [8]. Only one industrial time-series paper gives a concrete coverage and width number, 95.6% coverage and a 15.70% width reduction versus a benchmark, but this is on a sintering production dataset, not financial data [5]. A builder should therefore treat this note as a specification for an experiment still to be run: implement the PID or weighted-conformal wrapper described here, then measure coverage and width on OHLCV data directly, because the literature does not yet supply that number. Weighted conformal prediction is not specified for streaming price data in these sources either, so that adaptation is also left to the builder [1].
Fine-Tuning Time-Series Foundation Models on Intraday Bars: Which Recipe Beats Zero-Shot?
There is no direct evidence in the verified claims that answers this question. No cited passage compares full fine-tuning, LoRA, and quantile head replacement for TimesFM, Chronos, Lag-Llama, or Moirai on one-minute or five-minute OHLCV bars under purged walk-forward splits, and none reports GPU-hour costs for training a 25M-200M parameter model on a single asset with 500K-2M bars [1] [8]. What the claims do offer is background on how these foundation models are built and fine-tuned in other domains, general findings on why fine-tuning can help or hurt on unseen data, a parameter-efficient fine-tuning mechanism (LoRA) that could be tested, and a cautionary result from a related but different intraday forecasting study on futures data that found no statistically significant directional edge with sequential machine learning models. A builder who wants an answer must run the head-to-head experiment itself; this note describes what is known, what is missing, and how to design that experiment.
LLM Agent Planning: Does Hierarchical Decomposition Improve Web Task Success?
The evidence does not include a controlled, head-to-head test of hierarchical decomposition against flat planning on the same benchmark, so no claim directly proves that decomposition improves web task success in general. What the claims do show is that several systems built around decomposition (CoAct, WebAgent, RaDA, Region4Web, SkillWeaver, AdaPlanner) report gains over their own prior baselines or prior methods, each on its own benchmark and under its own conditions [2][5][6][7][8][3]. These reported gains are not comparable to each other because they use different benchmarks, different baselines and different task sets, so they cannot be added or averaged into a single number for 'decomposition'. Separately, one analysis argues that even where planning is decomposed into layers, low-level execution, not high-level planning, remains the main source of failure [1]. So the honest summary is: decomposition-based methods often beat their specific baselines, but the source claims do not isolate decomposition as the cause, and they do not show it removes the execution bottleneck.
The source of sam is closed until requested: write to support@qourat.com and say what you want it for. Investors and partners: the same address reaches the owner. All projects.