Home/Resources/Methodology
Resources

How we forecast the market

No black boxes. We believe you should understand how predictions are made before you trust them. Here's the methodology behind Nordict's forecasting engine.

Transparent approach, no hidden logic
Quantified uncertainty with confidence scores
Rigorous backtesting and validation
Honest about limitations and edge cases

Forecasting Pipeline

Simplified overview

1

Data Collection

Multi-source aggregation

2

Feature Engineering

Signal extraction

3

Model Ensemble

Multiple model voting

4

Confidence Scoring

Uncertainty quantification

5

Forecast Output

Direction + magnitude

Model overview

The core ideas behind our approach

You don't need a PhD to understand how we work. Here are the key concepts that power our forecasting engine, explained simply.

Ensemble Learning

We don't rely on a single model. Multiple specialized models vote on each forecast, reducing the risk of any one model's blind spots affecting predictions.

Simple analogy

Think of it like getting opinions from multiple experts rather than trusting just one.

Not a crystal ball

Our models are sophisticated, but they're not magic. They identify statistical patterns and quantify probabilities. Markets are inherently uncertain, and no model, ours included can predict the future with certainty. That's why confidence scores matter.

Data sources

What feeds our models

Good forecasts require good data. Here's the full picture of what goes into our models.

Price & Volume

Real-time updates

The foundation of any market analysis. We ingest high-frequency price and volume data across multiple timeframes.

Data points

OHLCV candles

1m to 1d resolution

Tick-level data

For volatility analysis

Volume profiles

By exchange and aggregate

Bid-ask spreads

Liquidity indicators

50+

Data sources

200+

Features extracted

5

Data categories

24/7

Data collection

Forecast process

From data to forecast in seconds

Here's exactly what happens when we generate a forecast—from raw data to actionable signal.

1

Data Ingestion

Collect & normalize

Raw data streams in from multiple sources—exchanges, on-chain providers, derivatives platforms. We normalize everything to a consistent format and timestamp.

Real-time

What happens here

Multi-exchange aggregation
Timestamp synchronization
Outlier detection & cleaning
Missing data interpolation

Validation

How we prove it works

Claims are easy. We hold ourselves to rigorous validation standards so you can trust the forecasts.

Historical Backtesting

We test our models against years of historical data, simulating how forecasts would have performed in real market conditions.

Methodology

Walk-forward analysis with expanding training windows. No future data leakage, models only see data available at forecast time.

Key metrics

2019–2025

Test period

2M+

Data points

50+

Assets tested

Want to see the numbers?

Our Performance page shows live accuracy metrics, historical backtests, and confidence calibration data.

View performance

Limitations

What we can't predict

Honesty about limitations is as important as confidence in strengths. Here's where our models fall short.

Our models learn from historical patterns. Events without precedent, exchange hacks, regulatory surprises, geopolitical shocks, can't be predicted from past data.

Example

The FTX collapse, Terra/Luna crash, or sudden regulatory bans are examples of events no model could have predicted in advance.

Sometimes our models simply don't know. During regime transitions or unusual market conditions, confidence scores drop. Low confidence means low predictability.

Example

When confidence is below 50%, the forecast is barely better than a coin flip. We show this clearly, don't ignore it.

Our shortest horizon is 4 hours. We don't attempt to predict minute-by-minute moves, that's dominated by noise, not signal.

Example

If you're scalping or trading on 1-minute charts, our forecasts won't help. They're designed for 4h+ decision-making.

Illiquid assets with thin order books behave differently. A single large order can move price 10% no model predicts that.

Example

We focus on assets with sufficient liquidity and trading history. New tokens or micro-caps aren't in our coverage.

Coordinated pump-and-dumps, wash trading, or spoofing create artificial patterns. Our models can be fooled by bad actors, just like human traders.

Example

If a whale group coordinates a pump on Telegram, our model sees bullish signals, but those signals are artificial.

Our models find statistical patterns, not causal relationships. A pattern that worked for 5 years can stop working without warning if the underlying dynamics change.

Example

Past performance doesn't guarantee future results. This isn't a disclaimer, it's a fundamental truth of statistical modeling.

Why we share this

Being upfront about limitations isn't weakness. It's how we build trust. If we only showed you the wins, you wouldn't have the full picture. Use our forecasts as one input among many, not as the sole basis for decisions.

See it in action

Now that you understand how it works, try it yourself. Start with a 7-day free trial of Premium.