Multimodal Market AI — synthetic demo

Deterministic signals and a visual signal from a chart image are fused into one row per decision. Everything on this page is fake data, generated by open-source code.

Not a trading bot, not a signal service, not a promise of profit. The visual signal here is a deliberately simple mock, used to show how an image path, a visual feature and a fusion record fit together.

The pipeline

synthetic OHLCV
      |
      +------> deterministic signals -----+
      |                                   |
      +------> chart image -> mock visual-+
                                          |
                                          v
                                     fusion table
                                          |
                                          v
                                  LONG / WAIT / SHORT

Fusion table

One row per decision. Every input already existed at the decision time; future outcomes are never used as inputs.

Case	Decision time (UTC)	Trend	Momentum	Structure	Visual score (mock)	Visual signal	Deterministic	Decision
case_001	2026-01-02 05:00	+1	+1	+1	+0.0188	+1	+1	LONG
case_002	2026-01-03 11:00	0	0	-1	-0.0024	0	-1	WAIT
case_003	2026-01-04 17:00	-1	-1	-1	-0.0188	-1	-1	SHORT

Swipe the table sideways to see every column, including the final decision.

Reading the signals: +1 means a rule points one way, −1 the other way, and 0 means that rule is not voting. A final LONG or SHORT needs the deterministic result and the visual signal to agree; otherwise the decision is WAIT.

The three chart images

Synthetic price line rising over 20 bars — case_001 — up-trending teaching case

Synthetic price line moving sideways with small swings — case_002 — mixed teaching case

Synthetic price line falling over 20 bars — case_003 — down-trending teaching case

The mock visual adapter

In a real setup, a vision model would look at each chart image and answer with a direction. Here a small stand-in plays that part, so the demo runs anywhere with no model and no GPU. It does not read the image: it measures the price change over the last 12 closes (mock_visual_score) and turns it into +1 above +0.3%, −1 below −0.3%, and 0 in between. A real model would fill the same visual_signal column. It would first have to show that it adds information over a plain deterministic baseline.

Why it is built this way

Each row has a stable identity and a decision time.
Unavailable data is not invented.
Deterministic, numerical and visual sources stay separate enough to audit.
A visual model is optional: it has to prove it adds information over a plain baseline.

Latest public research snapshot

An October 2026 large-sample arena evaluated 5 open VLM configurations × 1,000 frozen cases, producing 5,000/5,000 valid outputs across two 500-case tranches per model.

Qwen3-VL 8B had the strongest Task A agreement with the deterministic comparison reference: 0.603.
Task B remained difficult for every candidate; the highest observed agreement was 0.496.
The comparison reference is an anchor, not ground truth, and the private task meaning remains closed.
The two tranches were stable and an independent recomputation found zero discrepancies in the published aggregates.

Read the full public-safe arena report.

Get the data and the code

Source code on GitHub (Apache-2.0)
Synthetic dataset on Hugging Face
Raw files for this page: fusion_table.csv, market.csv