Which AI calls market direction best?

AI Forecasting Research · Updated Oct 27, 2026

Short answer. Among the 9 models with 10 or more graded calls, Grok called direction right most often: 58% (37 of 64). The lowest was ChatGPT — GPT-6, at 32% (16 of 50); 2 of the 9 did better than a coin flip. 3 models have fewer than 10 graded calls, an insufficient sample, and are left out of this comparison.

What this measures. For each model, the share of its directional calls (a forecast range pointing up or down from the starting price) that were right once the window closed, and how often the close landed inside its range. Each author's forecasts are counted on their own and never pooled with another's: two people choosing the same model are not running the same experiment.

Recorded by Forecastnest

928 public forecasts, 928 settled. Every model here answered this author's prompts, so the comparison holds everything but the model fixed.

Direction right, by model — Forecastnest

Full figures

Model	Forecasts	Settled	Directional calls graded	Direction right	Close inside the range	Average score
DeepSeek	144	144	103	41% (42 of 103)	78% (112 of 144)	48.5
Muse Spark	148	148	99	45% (45 of 99)	78% (116 of 148)	51.4
Gemini — Pro	139	139	78	55% (43 of 78)	70% (97 of 139)	52.8
Perplexity	83	83	67	45% (30 of 67)	82% (68 of 83)	50.5
ChatGPT — GPT-5	84	84	66	42% (28 of 66)	86% (72 of 84)	50.1
Grok	134	134	64	58% (37 of 64)	82% (110 of 134)	53.3
Claude — Opus	76	76	52	37% (19 of 52)	80% (61 of 76)	42.6
ChatGPT — GPT-6	60	60	50	32% (16 of 50)	90% (54 of 60)	49.7
Claude — Sonnet	44	44	19	37% (7 of 19)	82% (36 of 44)	48.4
Gemini — Flash (insufficient sample)	7	7	5	80% (4 of 5)	100% (7 of 7)	66.9
Kimi (insufficient sample)	5	5	4	75% (3 of 4)	100% (5 of 5)	68.9
Other model (insufficient sample)	4	4	2	50% (1 of 2)	75% (3 of 4)	49.7

By horizon

Model	Horizon	Directional calls graded	Direction right	Close inside the range
DeepSeek	4 hours	1	0% (0 of 1)	0% (0 of 1)
DeepSeek	1 day	101	41% (41 of 101)	78% (111 of 142)
DeepSeek	1 week	1	100% (1 of 1)	100% (1 of 1)
Muse Spark	1 day	99	45% (45 of 99)	78% (116 of 148)
Gemini — Pro	4 hours	1	100% (1 of 1)	0% (0 of 1)
Gemini — Pro	1 day	77	55% (42 of 77)	70% (96 of 137)
Gemini — Pro	1 week	0	—	100% (1 of 1)
Perplexity	1 day	67	45% (30 of 67)	82% (68 of 83)
ChatGPT — GPT-5	4 hours	1	0% (0 of 1)	0% (0 of 1)
ChatGPT — GPT-5	1 day	64	42% (27 of 64)	87% (71 of 82)
ChatGPT — GPT-5	1 week	1	100% (1 of 1)	100% (1 of 1)
Grok	4 hours	0	—	0% (0 of 1)
Grok	1 day	64	58% (37 of 64)	83% (109 of 132)
Grok	1 week	0	—	100% (1 of 1)
Claude — Opus	4 hours	1	100% (1 of 1)	0% (0 of 1)
Claude — Opus	1 day	51	35% (18 of 51)	81% (61 of 75)
ChatGPT — GPT-6	1 day	50	32% (16 of 50)	90% (54 of 60)
Claude — Sonnet	1 day	19	37% (7 of 19)	82% (36 of 44)
Gemini — Flash	1 day	5	80% (4 of 5)	100% (7 of 7)
Kimi	1 day	4	75% (3 of 4)	100% (5 of 5)
Other model	1 day	2	50% (1 of 2)	75% (3 of 4)

Limitations

• Most forecasts so far look one day ahead, on 9 assets: these rates describe that task, not market forecasting in general.
• A model is recorded from its own answer and checked against the other answers to the same prompt; an answer found filed under the wrong model is corrected by hand.
• A record of fewer than 10 graded calls is an insufficient sample: one forecast can move it a long way.

Citation and data provenance

Source

ForecastNest public forecasting dataset: forecasts their author chose to publish

Observation period

Aug 25, 2026 – Oct 20, 2026

Counted from

Aug 24, 2026

Public forecasts counted

928, of which 928 settled

Published but not counted

29, made before the baseline

Excluded for look-ahead

192 stock and ETF forecasts made outside market hours against the previous close, after news that close did not reflect

Methodology

How scoring works

License

CC BY-NC 4.0: non-commercial reuse with credit to ForecastNest and a link to this page

Underlying records

Every public forecast, each on its own page

Last updated

Oct 27, 2026

← All research questions