Home/Product/Forecasting Engine
Deep Dive

The forecasting engine behind the predictions

A closer look at how Nordict generates forecasts from multi-horizon predictions and confidence scoring to the model lifecycle that keeps everything accountable.

4

Time horizons

95%

Confidence bands

24/7

Forecast updates

Market Data Input

Real-time & historical

N
Forecasting Engine
v2.4.1
Feature Eng.
ML Models
Calibration

Forecast

+2.4%

24h horizon

Confidence

72%

Calibrated

Supported horizons

Choose the time horizon that fits your strategy.

Different decisions need different timeframes. Nordict supports multiple horizons, each with appropriate confidence calibration.

24 Hours

24H

Short-term forecasts for capturing near-term price movements. Updated hourly with tight confidence intervals.

Best for

Day tradersScalping strategiesTiming entries/exitsHigh-frequency decisions

Update frequency

Hourly

Confidence precision

High

Signal volatility

Higher

Horizon comparison

24 Hours
24H
30 Days
30D
12 Weeks
12W
12 Months
12M

Confidence bands & probabilities

Understanding probability ranges, not just predictions.

Confidence bands show where price might land with different probabilities. They turn uncertainty into something you can plan around.

Confidence Band Visualization

Hover over bands to explore

BTC 24h
+10%+5%0%-5%
Now+24h
Hover over a band to see details
50%
75%
95%

Wider bands = more uncertainty

When the model is less certain, bands expand. This isn't a flaw, it's honest communication about forecast reliability.

Calibration matters

A 75% band should contain the actual outcome ~75% of the time. We test this continuously and adjust when needed.

Not trading signals

Bands show probability, not recommendations. A narrow band doesn't mean 'trade now', it means the model is more certain.

Regime-aware sizing

During volatile periods, bands widen automatically. During calm periods, they tighten. The model adapts to conditions.

Model lifecycle

From training to live monitoring, every step tracked.

Models go through a rigorous lifecycle before generating forecasts you see. Here's how it works.

01

Training

Models are trained on historical market data using walk-forward methodology. No future data ever leaks into training.

Key aspects

1Historical data spanning multiple market regimes
2Feature engineering with domain expertise
3Strict temporal data separation
4Multiple model architectures evaluated

This cycle repeats continuously. When monitoring detects drift, retraining is triggered automatically.

Update frequency & data freshness

Fresh data, appropriate cadence.

Different horizons need different update frequencies. More frequent isn't always better. It's about matching cadence to decision timescale.

Forecast Update Schedule

20:20:42 UTC

24 Hours

Every 6 hours

Refreshed four times daily to capture short-term market movements.

Data lag< 5 min
Last updateNext update

30 Days

Daily

Updated once per day to balance responsiveness with stability.

Data lag< 1 hour
Last updateNext update

12 Weeks

Daily

Updated once per day. Longer horizons benefit from stable daily updates.

Data lag< 1 hour
Last updateNext update

12 Months

Daily

Updated once per day for long-term strategic forecasts with maximum stability.

Data lag< 1 hour
Last updateNext update

Data Source Freshness

Price Data

Major exchanges

Real-time

Volume Data

Aggregated feeds

Real-time

Order Book

Top 5 exchanges

Near real-time

On-chain (Crypto)

Node providers

~10 min lag

Why some data is delayed

On-chain data requires block confirmations for accuracy. We prioritize correctness over speed for data that influences longer-horizon forecasts.

99.7%

Data pipeline uptime

<30s

Avg. processing time