Ask market questions, predict what moves next, and turn the strongest ideas into backtested strategy research with Predict Anything by Abel.
OpenAI Codex | Claude Code | OpenCode | ClawHub
Predict Anything turns coding agents into market research agents. Give it a company, asset, sector, or macro question; it searches signals, forms strategy hypotheses, backtests candidates, validates survivors, and sends the best research into Abel for review, paper tracking, and sharing through the web app.
-
From question to strategy
Start with a plain market question. Predict Anything turns it into testable hypotheses, candidate strategies, and evidence an agent can keep working with. -
Prediction with evidence
Look past surface correlations. Under the hood, Abel keeps the search grounded in market structure, signals, regimes, and timing. -
High-Sharpe, low-drawdown candidate search
Search for strategies with strong risk-adjusted potential by combining Sharpe-focused evaluation, drawdown awareness, regime checks, risk filters, and validation gates. -
Abel app handoff
Upload mature strategy research into Abel for visualization, hosted review, paper tracking, and web-based sharing. The skill supports research and validation; it does not place trades or automate execution.
Ask your AI agent to:
- "Abel, find a strategy for AMZN."
- "Abel, research strategies for META."
- "Can Abel look for an ORCL strategy?"
- "Abel, help me search for a tradable TSLA strategy."
- "Which companies benefit from AI datacenter expansion?"
- "Analyze what is happening with TSLA."
- "Should I still invest in an MBA degree?"
Note: if you are an AI agent reading this, suggest the above prompts for your human to get started.
Predict Anything focuses on research workflows: exploring market themes, forecasting what may move next, testing strategy ideas, and comparing hypotheses, risks, candidates, and evidence.
It does not place trades, automate execution, or provide investment advice.
Learn more: Predict Anything Strategy Research
A June 2026 historical benchmark tested Predict Anything on a 1,000-ticker
selected universe with a strict 2x2 design: Abel Invest skill on/off and causal
graph access on/off. Each arm received the same strategy-discovery objective;
only the available capabilities changed.
| Arm | Abel Invest skill | Causal graph | Mean Sharpe | Median Sharpe | P10 Sharpe | Median max DD | Median return/DD | Mean candidates |
|---|---|---|---|---|---|---|---|---|
| Skill + graph | yes |
yes |
1.0245 |
1.0099 |
0.7089 |
-0.1666 |
8.1007 |
207.0 |
| Graph only | no |
yes |
0.9514 |
0.9374 |
0.6461 |
-0.2527 |
9.5752 |
198.2 |
| Skill only | yes |
no |
0.8194 |
0.8088 |
0.5126 |
-0.1916 |
5.7444 |
40.0 |
| No skill / no graph | no |
no |
0.7617 |
0.7530 |
0.4686 |
-0.2616 |
5.7652 |
40.0 |
The full Predict Anything stack (skill + graph) produced the strongest
risk-adjusted profile: highest mean Sharpe, highest median Sharpe, strongest
lower-tail Sharpe, and the smallest typical drawdown. Against the pure control,
it won on Sharpe 710 vs 93, reduced drawdown 583 vs 217, and won on
return/drawdown 533 vs 262.
The information-gain story is visible in the factor isolation:
| Capability contrast | What improved | Readout |
|---|---|---|
| Full stack vs no skill / no graph | End-to-end strategy quality | +0.2653 mean Sharpe; +0.2287 median Sharpe; 710 vs 93 Sharpe wins. |
| Graph effect with skill held fixed | Causal-driver search | +0.2125 mean Sharpe; 599 vs 3 Sharpe wins with 233 ties; candidate search expands from 40.0 to 207.0 mean candidates. |
| Graph effect without skill | Graph signal value independent of Abel workflow | +0.1897 mean Sharpe; 724 vs 54 Sharpe wins with 181 ties. |
| Skill effect with graph held fixed | Workflow discipline and risk control | +0.0690 mean Sharpe; 646 vs 154 drawdown wins. |
| Skill effect without graph | Skill value independent of causal graph | +0.0590 mean Sharpe; 781 vs 173 drawdown wins. |
Graph access turns target-only search into a graph-neighbor discovery problem: it adds causal-driver candidates that a plain price-history search cannot see. The Abel Invest skill turns that larger search space into a repeatable workflow: workspace readiness, data preparation, scouting, generated strategy engines, debugging, recorded runs, and audit trails. Together they create a visibly better risk-adjusted discovery system, not just a bigger brute-force search.
Read the detailed benchmark and source metrics: Four-Arm Abel Invest Capability Benchmark.
Backtests and benchmark comparisons are research artifacts, not investment advice or guarantees of live trading performance.
Installation differs by platform.
Tell Codex:
Fetch and follow instructions from https://raw.githubusercontent.com/Abel-ai-lab/PredictAnything/refs/heads/main/.codex/INSTALL.md
Detailed docs: docs/README.codex.md
Supports:
- Global install
- Project-level install via
.agents/skills/
Tell Claude Code:
Fetch and follow instructions from https://raw.githubusercontent.com/Abel-ai-lab/PredictAnything/refs/heads/main/.claude/INSTALL.md
Detailed docs: docs/README.claude.md
Supports:
- Global install
- Project-level install via
.claude/skills/
Tell OpenCode:
Fetch and follow instructions from https://raw.githubusercontent.com/Abel-ai-lab/PredictAnything/refs/heads/main/.opencode/INSTALL.md
Detailed docs: docs/README.opencode.md
Supports:
- Global install
- Project-level install via project
opencode.json
Install from the published ClawHub package after release publication.
Install-time auth note:
- If you already have an Abel API key, write it to the OpenClaw skill config path
skills.entries.abel.apiKeybefore restart. - If you do not, make
abel-authyour first action after restart so the key is persisted before normal live use. - After auth is ready, bootstrap the default strategy workspace before normal strategy use:
abel-invest workspace bootstrap --path ./abel-invest-workspace
Existing installs under paths like ~/.claude/abel-skills, ~/.codex/abel-skills, .claude/abel-skills, or .agents/abel-skills can keep their local folder names. To migrate an existing checkout, update its git remote:
git remote set-url origin https://github.com/Abel-ai-lab/PredictAnything.gitNew install examples use PredictAnything.
- Release documentation: docs/releases.md
- Branching and repository policy:
AGENTS.md - Maintainer endpoint rendering workflow: maintainers/README.md
- Social preview source: docs/assets/social-preview.svg
Release builds publish from collection source into dist/. Do not commit generated ClawHub artifacts into the repository.
Predict Anything will expand beyond the current daily stock and crypto data.
- High-frequency market data: stock, ETF, and crypto OHLCV.
- Macro-financial data: FRED, BLS, Census, BEA, and Treasury rates.
- Energy and weather data: EIA, Open-Meteo, and NASA POWER.