Temporal AI Research & Products

Proof, not promises.

EventHorizon-AI builds forecasting systems for crypto markets and retail demand, and puts every one of them through the same statistical validation pipeline before it ships. No cherry-picked backtests — walk-forward, bootstrap, and permutation testing, published openly.

validation.log
eventhorizon validate --model crypto-btc-5s
walk-forward  (temporal embargo)PASS
block bootstrap  (per-date)PASS
paired gap bootstrap  (vs. baseline)PASS
permutation testPASS
edge+8.0pp  (60.1% vs 52.0%)
4-stage statistical validation on every product
Validation library open-sourced
Limitations documented, not buried
Products

Two systems. One pipeline.

Each product below is validated with the same methodology described further down this page — not a separate marketing claim per product.

EventHorizon Crypto
BTC/USDT · 5-second horizon
Validated
+8pp
edge over baseline — 60% vs 52% directional accuracy
⚠️ Honest limitation: economically unviable under standard exchange fees at this horizon. Documented openly — see the audit dashboard for the full breakdown.
View audit dashboard →
EventHorizon Demand
Retail demand · M5 (Walmart) dataset
Validated
30%
WAPE improvement — 30.9% vs 44.4% seasonal baseline
ℹ️ Status: statistically significant in 7 of 7 stores tested (per-date bootstrap, paired gap vs. baseline). Currently accepting pilot partners.
Apply for pilot →
Methodology

The validation pipeline

Every product above runs through these four checks, in this order, before a single number is published.

01 / walk-forward

Temporal embargo

The model never sees data from around the prediction boundary — no leakage from the future into training.

02 / block bootstrap

Per-date resampling

Confidence intervals are built by resampling contiguous date blocks, respecting the autocorrelation in time series.

03 / paired gap bootstrap

Model vs. baseline

The gap between model and baseline is bootstrapped directly, in pairs, instead of comparing two separate confidence intervals.

04 / permutation test

Null-hypothesis check

Labels are shuffled and the model is re-scored, to confirm the observed edge isn't a product of chance.

The full methodology is packaged as an installable Python library: honest-validation-toolkit

View on GitHub →
Get in touch

Interested in a pilot?

If you run a small or medium retail business and want to try EventHorizon Demand for free, send an email to lucasthd08@gmail.com with the subject "Pilot Demand" and attach a CSV of your daily sales. I'll send you a forecast within 24 hours — no cost, no commitment.

Send email →

or message directly:

Send message via WhatsApp →

Prefer GitHub? Reach out there — issues, discussions, or a direct message all work.