# Analysis Gym

Analysis Gym is a tiny MCP server for recording and scoring prospective equity
earnings predictions made by AI agents.

It deliberately does not choose tickers, schedule runs, or invoke models. Your
agent loop owns those decisions. The agent uses the existing
[FactIQ](https://factiq.com) MCP server for research and calls Analysis Gym only
to record a prediction, record the eventual actuals, or read the results.

## Tools

- `record_prediction` records an immutable forecast before the expected
  earnings time.
- `record_actuals` settles all earlier predictions for a ticker and fiscal
  period.
- `get_results` returns per-metric errors and a leaderboard grouped by harness,
  model, and thinking setting.

The five predicted values are revenue, EBITDA, net profit, free cash flow, and
the first regular-session closing price after the earnings release.

## Run locally

```bash
uv sync
uv run analysis-gym
```

The server uses stdio transport and stores data in `analysis_gym.sqlite3` in its
working directory. Set `ANALYSIS_GYM_DB_PATH` to put the database elsewhere.

### Codex

Add the server to `~/.codex/config.toml`:

```toml
[mcp_servers.analysis-gym]
command = "uv"
args = ["--directory", "/absolute/path/to/analysis-gym", "run", "analysis-gym"]
```

Install and authenticate the FactIQ plugin separately. Then ask Codex, for
example:

> Pick an equity reporting soon. Use FactIQ to forecast its next-quarter
> revenue, EBITDA, net profit, free cash flow, and first post-earnings close.
> Record the forecast in Analysis Gym before the release.

### Claude Code

```bash
claude mcp add analysis-gym -- \
  uv --directory /absolute/path/to/analysis-gym run analysis-gym
```

Use the same prompt and ensure the FactIQ plugin is also installed and
authenticated.

## Agent-side loop

A loop outside this repository can choose an upcoming event and run the same
request through any set of CLI/model/thinking configurations. Each agent calls
`record_prediction` itself. After earnings, call `record_actuals` once with a
source URL, then use `get_results` to compare the configurations.

Analysis Gym uses symmetric mean absolute percentage error (SMAPE), where lower
is better. It reports every metric separately and a simple mean across all five.

## Metric definitions

- EBITDA: operating income plus depreciation and amortization.
- Free cash flow: operating cash flow minus capital expenditure.
- Net profit: consolidated net income attributable to the parent/common
  shareholders.
- Post-earnings close: the same session's close for a pre-market release, or the
  next regular session's close for an after-hours release.

All four financial values (submitted in millions) in a submission must use the same reporting currency.

## Development

```bash
uv run pytest
```