IrwanGantenx/riak-backtests

Riak Backtests v0.3

Historical-scenario backtest cases for Riak, a self-contained (zero-dependency, deterministic rule-based) causal prediction engine.

Each case is a real historical episode: a short seed_text describing the situation as it looked at the time, plus expect_any — keyword alternatives for outcomes that actually materialised. The engine expands the seed into a cause→effect web and is scored on whether its top predictions match reality.

This is v0.3: 15 cases, English. v0.1 shipped 5 cases; v0.2 added 5 episodes in new domains (energy, lockdown, strike, banking, cyberattack); v0.3 adds 5 generalization variants inside covered domains (hospital crunch, port strike, rate shock, bank contagion, pension protests) to probe whether the engine generalizes rather than memorizing the first 10. Baseline comparisons (LLM-only reasoning, human forecasts) are planned and will land as versioned updates.

Current scores on engine main branch (commit 88f3fd5): hit@6 15/15, mean Brier 0.500. History: 1/5 before the top_outcomes ranking fix (commit 18e0a97), 5/5 after the fix (40bc109), 10/10 at v0.2. All numbers are reported so each improvement reads as a documented change, not a tuned result. Fifteen author-labelled cases remain a small sample, and Brier near 0.5 means probabilities are uncalibrated; no general accuracy claim is made.

Selection honesty note: candidate seeds were probed pre-commit and kept only if they yield a rich web (12+ nodes); rejected candidates are listed in the GitHub PR messages (PR #3, PR #5). This favours the engine by construction, so the upcoming LLM baseline on the same cases is the fair comparison.

Related artefacts: interactive demo (precomputed outputs, runs in the browser) · scored evaluation notebook (reruns the backtest on free hardware) · interactive explorer notebook (run your own scenario).

Format

data/backtests/*.json:

{
  "name": "COVID-19 outbreak 2020",
  "seed_text": "A novel virus outbreak spreads across cities...",
  "lang": "en",
  "seed": 2020,
  "expect_any": ["healthcare demand", "preventive measures", "supply chain", "anxiety"]
}

Case	Episode	Seed
covid-outbreak-2020	COVID-19 outbreak	2020
fuel-price-hike-2022	2022 fuel price hike	2022
social-media-regulation	Social media regulation wave	2021
ai-automation-wave	AI-driven automation	2023
deforestation-pressure	Deforestation pressure	2019
hospital-crunch-2021	Hospital overload wave	2021
port-strike-2023	Container port shutdown	2023
rate-shock-2022	Overnight rate doubling	2022
regional-bank-contagion-2023	Regional bank runs	2023
pension-reform-protests-2023	Retirement age protests	2023
bank-panic-2023	Single bank failure	2023
energy-crunch-europe-2022	European gas cutoff	2022
pipeline-cyberattack-2021	Fuel pipeline shutdown	2021
rail-strike-2022	Nationwide rail strike	2022
shanghai-lockdown-2022	City-wide lockdown	2022

Reproduce the scores

backtest.py (requires the Riak engine source) runs all cases offline and reports two metrics:

• hit@6 — does any of the top-6 predicted outcomes match an expected keyword?
• Brier — (1 − p)² where p is the probability of the best-matching node (0 = perfect, lower is better).

git clone https://github.com/irwanformal-cmd/riak
cd riak
cp /path/to/this/dataset/data/backtests/*.json backtests/
python3 backtest.py

A runnable version of this evaluation also exists as a Kaggle Notebook so anyone can re-run it on free hardware.

Limitations

• 5 cases is a smoke test, not a benchmark. No claim of general predictive accuracy is made.
• expect_any keywords are author-labelled, not independently annotated.
• Deterministic rule-based mode only; LLM-augmented mode is not covered here.

License & citation

MIT. If you use this dataset, cite Riak via CITATION.cff (also in this repo).