Polarıs

01 / Introduction

Introducing MFM-1 Polaris

Polaris is the first model in the llmcharts Market Foundation Model family — a dedicated stack for understanding what candlestick markets may do next. Where general-purpose language models reason in words, Polaris is trained and served around K-line structure: open, high, low, close, and volume across time.

Markets are always moving. Traders already stitch together indicators, narratives, and intuition to form a view. Polaris begins our work on a foundation that can ingest those same series and return structured forecast paths — mean trajectories and, when sampled, distributions over possible futures — without pretending to know what “should” happen.

This page is the product introduction for generation 01. It explains how we think about the problem, how we intend to evaluate models like Polaris, and how you can access the three power modes from the llmcharts terminal or a public HTTP API.

02 / Approach

From patterns to possibilities

A Market Foundation Model starts from observation. Given a window of recent OHLCV bars, Polaris produces a forward path of predicted candles for a chosen horizon. The input contract follows the practical needs of charting: intervals from minutes to days, lookbacks sized to each tier’s context window, and optional sampling when you want a band rather than a single line.

Polaris is a market-sequence transformer stack hosted and productized by llmcharts as mini, medium, and high power modes. Those modes trade latency and fidelity the way language-model tiers do, but the “tokens” are market bars, not chat turns.

Polaris is designed to live in two places at once: inside the llmcharts terminal as an overlay and assistant tool, and outside the product as an API you can call the way you call any other foundation-model endpoint. The same model IDs and request shape apply in both contexts.

What Polaris is not

Polaris does not place orders, manage risk for you, or issue investment advice. Forecast paths are informational scenarios — useful for research, alerts, and chart narration — not a signal to trade. Outputs can be wrong in quiet markets and wrong in chaotic ones; treat them as one more lens, never as a substitute for your own process.

03 / Evaluation

The evidence behind the model

Understanding a forecasting model means understanding where it works, how it is measured, and where its limits lie. mfm-eval is our evaluation suite for market foundation models — not chat arenas — with a path toward an open package (including Hugging Face) so results can be reproduced and compared.

Performance (mfm-eval-btc1h-v0)

Point metrics are MAE, RMSE, and sMAPE on close over a 24-bar horizon. Directional hit-rate is whether the mean path’s sign of return matches the realized move. Band coverage is the fraction of realized closes inside the returned low/high envelope (sample_count = 8). Lower MAE/RMSE/sMAPE is better; higher hit-rate and coverage is better.

Report: packages/mfm-eval/reports/mfm-eval-btc1h-v0.json · BTCUSDT 1h · 48 non-overlapping origins · lookback 128 · pred 24 · model_version 2026.09.24-mfm · run 2026-09-24 · runner runpod. Scores are from that artifact only — never invented.

*Last-close has no directional call (flat forecast). On this slice, naïve last-close still leads on point error — Polaris is not claimed to beat every baseline on every window. Hit-rate and band coverage are the distributional story; expand to more regimes in later mfm-eval releases.

Methodology

Fixed held-out windows with non-overlapping prediction horizons; zero-shot hosted weights (no fine-tune overlap claim). Crypto first: BTCUSDT 1h for v0. Each power mode is reported separately. Baselines: last-close naïve, drift, seasonal naïve (−24h), and AR(1) on close returns. Reproduce with mfm-eval run --dataset btc1h-v0 --runner runpod.

Scope & limitations

Generation 01 focuses on liquid crypto pairs and standard Binance-style intervals. Context windows differ by tier (mini supports longer lookbacks; medium and high prioritize fidelity within a tighter window). Latency and availability depend on hosted GPU capacity. Always read the disclaimer returned with every forecast response.

04 / Access

Meet your next point of reference

Call Polaris from the llmcharts platform or via POST /v1/forecast with an API key. Model IDs are versioned with the family and generation:

Pricing

Usage is metered on compute: lookback bars plus predicted bars, times sample count, times a tier multiplier. A per-request floor covers sparse GPU idle cost on the hosted fleet.

units    = (lookback + pred_len) × max(1, sample_count)
variable = units × $0.000004 × tier_multiplier   # mini 1× · medium 2.5× · high 5×
bill     = max($0.04, variable)

API shape

curl -s https://dev.llmcharts.com/v1/forecast \
  -H "Authorization: Bearer $POLARIS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"mfm-1-polaris-medium","symbol":"BTCUSDT","interval":"1h","pred_len":24}'

Informational model output only — not investment advice. LLMCHARTS does not execute trades from Polaris forecasts.