How accurate are endeavr.ai's machine-learning stock predictions?

endeavr.ai publishes live ensemble accuracy per horizon on the Accuracy dashboard. Ensemble 1-day directional accuracy averages in the high 50s on large-cap US equities across full market cycles. Every prediction ships with a calibrated confidence interval. We do not hide underperformance.

What deep-learning architectures power endeavr.ai's stock predictions?

endeavr.ai uses a four-model neural ensemble: LSTM with attention, GRU with skip connections, Temporal CNN in WaveNet style, and a Transformer encoder. Each model is trained on 45 engineered features spanning price, volume, volatility, and macro regime factors. Model weights are versioned, audited, and walk-forward backtested before promotion.

How often are the ML models retrained?

Per-ticker fine-tuning runs nightly on the most recent rolling window. Full architecture retraining with hyperparameter search runs weekly. Every retrain is benchmarked against the prior weights on a held-out walk-forward set before promotion to production.

Where does endeavr.ai source its market data?

endeavr.ai aggregates data from multiple licensed providers including Yahoo Finance fundamentals, Polygon equities, and curated news feeds. Real-time quotes stream through our edge broadcast hub.

How much does endeavr.ai cost?

endeavr.ai is free to start with 5 full AI analyses per day. Pro is $19.99/month (or $199.99/year) with unlimited analyses, ML predictions, backtesting, and risk analytics. Workspace is $49.99/month for teams and includes API access. Every plan includes a 7-day Pro trial with no credit card required.

Is endeavr.ai investment advice?

No. endeavr.ai is a research and analytics platform for informational and educational purposes only. We do not provide personalized investment advice. Please consult a licensed advisor before making investment decisions.

Does endeavr.ai offer an API?

Yes. API access and webhook exports are included on the Workspace tier. Endpoints cover predictions, quantitative metrics, and daily intelligence briefs. Rate limits are generous for research workflows.

How is endeavr.ai different from Bloomberg Terminal or QuantConnect?

Bloomberg Terminal costs around $24,000 per year and targets institutional desks. Quant data vendors run $5,000 to $15,000 per year. endeavr.ai delivers a production deep-learning ensemble plus institutional risk analytics for $19.99 per month — roughly 99% cheaper than Bloomberg with modern UX and out-of-the-box ML.

What is walk-forward backtesting?

Walk-forward backtesting evaluates a model by training on a rolling historical window, predicting the next out-of-sample period, then sliding the window forward. It prevents look-ahead bias and is the institutional standard for ML model validation. endeavr.ai walk-forward backtests every model nightly.

Can I cancel my endeavr.ai subscription at any time?

Yes. Cancellations take effect at the end of the current billing period. You retain full access until then and we never charge surprise fees.

Inside endeavr.ai's Four-Model Deep-Learning Ensemble

2026-03-09 · endeavr.ai Research

There is no single best neural architecture for financial time series. Each captures different structure and fails in different ways. That is exactly why we use four of them and stack their predictions rather than betting the platform on one.

The four base models

An LSTM with attention captures momentum and mean-reversion over weeks. A GRU with skip connections offers a lighter recurrent view that trains faster and resists overfitting. A Temporal CNN in the WaveNet style reads multi-scale patterns through dilated convolutions. A Transformer encoder relates any two days in the lookback window directly through self-attention.

Why stacking beats any single model

Because the four models make uncorrelated errors, a stacking meta-learner that combines them cancels idiosyncratic mistakes and produces a more robust forecast than the best individual model. It is the same reason a panel of independent forecasters beats any single pundit.

Calibrated, not just confident

Raw model probabilities are usually poorly calibrated. We calibrate ensemble confidence on held-out data so that a 70%-confidence prediction is genuinely right about 70% of the time — which is what makes conviction-weighted position sizing meaningful.