INDEPENDENT AI & MACHINE LEARNING ENGINEERING

Forecasting.
Validation.
Evidence.

Developing machine learning systems for time-series forecasting, alongside statistical tools to evaluate whether their results are reliable and useful.

Technical work. Honest validation. Evidence over promises.

FIG. 01RESEARCH SYSTEM / OVERVIEW
Illustrative time-series signal A conceptual line plot with a muted observed signal, a forecast segment, and a vertical evaluation boundary. This is not measured project data. OBSERVED SERIES FORECAST* evaluation boundary
* Conceptual illustration, not project dataSignal ≠ evidence
RESEARCH PRINCIPLEMeasure twice. Claim once.
01 — 05RESEARCH THAT SHOWS ITS WORK
TIME-SERIES FORECASTINGSTATISTICAL VALIDATIONREPRODUCIBLE MLOPEN-SOURCE ENGINEERING

01 / RESEARCH & SYSTEMS

Two problems.
One standard of evidence.

Each system asks a different forecasting question. Both are documented with the same distinction in mind: a model result is not automatically a practical outcome.

D/01RETAIL DEMAND
FORECAST ERROR / WAPEM5 · WALMART
≈30.1%relative WAPE reduction
vs. seasonal baseline

DOCUMENTED BENCHMARK RESULT · NOT CUSTOMER OUTCOMES

LIGHTGBMDAILY DEMAND FORECASTING
SYSTEM 01 Technical benchmark

EventHorizon Demand

Can machine learning improve daily retail demand forecasts over a seasonal baseline on a public benchmark?

31.03%Model WAPE
44.4%Seasonal baseline WAPE

What the result means: a ≈30.1% relative reduction in WAPE, calculated from 31.03% versus the 44.4% seasonal baseline, on the M5 Walmart dataset. The improvement was observed in all 7 stores evaluated, with statistical significance reported in the project documentation. This is a benchmark result: no commercial pilots have been executed, and practical value in real operations remains under investigation.

Approach
LightGBM trained on M5 Walmart daily sales, using sales history plus the temporal, rolling, price and promotion features available in the dataset.
Validation
Temporal holdout, date-block bootstrap and paired comparison against a seasonal baseline.
Limitations
Benchmark performance is technical evidence, not proof of lower waste, higher profit or customer impact. Commercial validation has not been performed.
C/02MARKET MICROSTRUCTURE
Conceptual short-horizon price path An illustrative zigzag price line showing why short-horizon directional prediction is challenging. It does not represent measured BTC/USDT prices. ILLUSTRATIVE SIGNAL · NOT MARKET DATA
BTC / USDT5-SECOND HORIZON
SYSTEM 02 Negative economic result

EventHorizon Crypto

Can very short-term price direction be predicted beyond a simple baseline — and does that signal survive execution costs?

≈60%Directional accuracy
≈52%Baseline reference
≈1.5MSamples evaluated

What the result means: an experimental directional edge was observed at the 5-second horizon. It did not translate into economic viability under the evaluated exchange-fee assumptions. This is not a claim of profitability or financial advantage.

Approach
Short-horizon BTC/USDT directional classification with a Transformer, BiLSTM and Cross-Attention architecture; temporal evaluation and execution simulation.
Validation
Walk-forward checks, block bootstrap, paired gap analysis and permutation testing are documented in the project audit.
Limitations
Checkpoint selection and execution assumptions constrain interpretation. Queue position, partial fills and adverse selection were not fully modeled.

Metrics are reported experimental results from the project documentation. Values are rounded for readability; methodology, dataset and limitations should be reviewed before drawing conclusions.

02 / CURRENT RESEARCH PHASE

From benchmark to
real-world validation.

The technical benchmarks are documented. The next challenge is understanding whether these methods solve meaningful problems in real operational environments.

EventHorizon is currently investigating practical use cases, evaluating real-world constraints and exploring where predictive systems can create measurable value.

No commercial pilots have been executed to date.

  • TECHNICAL BENCHMARKSCompleted
  • COMMERCIAL VALIDATIONOngoing

03 / VALIDATION ENGINEERING

A result is only
the beginning.

A single test metric does not establish generalization, reliability or practical usefulness. The validation process asks whether the result holds up over time, how uncertain it is, and whether it matters outside the metric.

STAGE 01 / 05 METHOD OVERVIEW

Start with a comparison.

A model score is hard to interpret in isolation. Compare it with a clearly defined baseline on aligned evaluation data before describing an improvement.

A better score is an observation — not, by itself, proof of generalization or business value.

THE PRINCIPLE

These methods help investigate specific failure modes. They do not automatically eliminate all bias or guarantee that a model will generalize.

04 / REUSABLE EVALUATION INFRASTRUCTURE

Honest Validation
Toolkit.

An open-source Python library for investigating time-series model performance. It makes comparisons and uncertainty explicit, helping reduce misleading conclusions about how well a model really performs. It is useful beyond any single EventHorizon system.

  • Block bootstrap
  • Gap bootstrap
  • Permutation tests
  • Walk-forward split
  • WAPE metric
  • MASE metric
Explore the toolkit
TOOLING / 001

Built to make evaluation more inspectable. It does not eliminate every validation problem, and no statistical test becomes a guarantee.

View license

05 / ABOUT

BRAZIL · INDEPENDENT

Curiosity is the start.
Evidence is the standard.

EventHorizon-AI is an independent research and engineering project developed in Brazil by Lucas.

Its work explores machine learning, time-series forecasting and statistical validation, with an emphasis on reproducibility, transparency and practical relevance. Negative results are documented as carefully as positive ones.

01 / Rigor over hype02 / Transparency over polish03 / Reproducibility over assumption

OPEN TECHNICAL WORK

Evidence over promises.

Explore the code, review the limitations and inspect the methods behind the results. Technical feedback, collaboration ideas and research discussion are welcome on GitHub.