Datasetpublic

WhatHui/fortivio-graded-stock-calls

Fortivio graded AI stock-prediction calls Every directional call made by the Fortivio.AI agent fleet on US-listed stocks, locked before the market open and graded after the close against that session's open→close move. Nothing is backtested and nothing is revised after the fact: a call is written once, before 09:30 ET, and scored once, mechanically, when the session's final price bar arrives. One row per graded call. One file per month. Monthly summaries with hit rate by… See the full description on the dataset page: https://huggingface.co/datasets/WhatHui/fortivio-graded-stock-calls.

sourceHugging Facecc-by-4.0updated 20d agoView on Hugging Face

1likes129downloads

Dataset Card

Fortivio graded AI stock-prediction calls

Every directional call made by the Fortivio.AI agent fleet on US-listed stocks, locked before the market open and graded after the close against that session's open→close move. Nothing is backtested and nothing is revised after the fact: a call is written once, before 09:30 ET, and scored once, mechanically, when the session's final price bar arrives.

One row per graded call. One file per month. Monthly summaries with hit rate by strategy, best and worst agents and tickers live at fortivio.ai/reports; the raw file for any month is https://fortivio.ai/reports/YYYY-MM.csv.

What's in it

month	graded calls	correct	hit rate	agents	tickers	sessions
2026-06 (from Jun 16)	12,686	6,365	50.2%	473	372	10
2026-07	81,852	40,122	49.0%	465	397	22
2026-08	62,944	31,984	50.8%	383	369	21

The fleet as a whole sits at a coin flip. That is the point of publishing it: the interesting structure is underneath — the spread between strategy families (LLM-prompt and MACD agents above 52% in August, low-volatility agents below 49%), specialist agents that hold 60–65% on their own domain over 80+ calls, and tickers the models read consistently well or consistently badly. The weak months are published with the same prominence as the strong ones.

Columns

column	type	meaning
`as_of_date`	date	the trading session the call was made for (the call day)
`agent_id`	string	stable agent id; profile and full history at `https://fortivio.ai/a/{agent_id}`
`agent_name`	string	display name
`strategy`	string	strategy family the agent runs (`momentum`, `macd`, `prompt`, `lightgbm`, `meta`, …)
`ticker`	string	US-listed symbol; history at `https://fortivio.ai/stock/{ticker}/predictions`
`direction`	`up` / `down`	the predicted open→close direction
`confidence`	float 0–1	the agent's self-reported confidence (may be empty)
`rank`	int	the stock's rank in the agent's cross-section that day (may be empty)
`correct`	0 / 1	whether the realized open→close direction matched
`realized_direction`	`up` / `down`	what the session actually did
`realized_return`	float	the session's open→close return (e.g. `-0.027059` = −2.71%)
`rationale`	string	the agent's one-line reason, generated at call time

Grading rules (short version)

—A call is accepted only before that session's open; after the open the prediction job refuses new calls for the day.
—Stored calls are insert-only and immutable.
—A call is graded against the same session's open-to-close return — the only move a pre-open call could have captured; the overnight gap is not counted.
—Grading uses final price bars only. If the call day's bar is missing the call stays ungraded and is not in this dataset.

Full methodology: https://fortivio.ai/methodology · glossary: https://fortivio.ai/glossary

Quick start

python

import pandas as pd
df = pd.read_csv("hf://datasets/WhatHui/fortivio-graded-stock-calls/data/graded-calls-2026-08.csv.gz")
print(df.groupby("strategy")["correct"].agg(["mean", "size"]).sort_values("mean", ascending=False))
print(df.groupby("direction")["correct"].mean())

python

from datasets import load_dataset
ds = load_dataset("WhatHui/fortivio-graded-stock-calls")

Live access for AI agents: Fortivio also hosts an MCP server at https://fortivio.ai/mcp (tools include get_monthly_report, get_stock_history, get_leaderboard).

Updates

A new file is added after each month closes. Closed months never change.

License and citation

CC BY 4.0. Please cite the monthly report page for the month you use, e.g.

Fortivio.AI, "AI stock-prediction accuracy — August 2026", https://fortivio.ai/reports/2026-08

Not investment advice. Model-generated predictions with a public graded record.