The Science Behind
the Forecast
Salmocast combines machine learning with domain expertise to deliver transparent, explainable salmon price forecasts. Here is how we turn raw data into actionable market intelligence.
48
Price Signals
13
Production Zones
4 Weeks
Forecast Horizon
Weekly
Model Updates
Our Approach
Quantitative rigor meets salmon market expertise
Salmon prices are driven by a complex interplay of biological cycles, regulatory decisions, global trade flows, and macroeconomic factors. Traditional forecasting methods struggle to capture these multi-dimensional relationships.
Our approach combines gradient boosting machine learning with deep domain knowledge of Norwegian aquaculture. We do not just fit curves - we model the fundamental supply and demand dynamics that move prices.
INPUT LAYER
FEATURE ENGINEERING
48 engineered features with lag transforms
LIGHTGBM ENSEMBLE
Gradient boosting with weekly retraining
OUTPUT
Data Pipeline
From raw data to actionable forecasts
Every Tuesday, our automated pipeline collects, processes, and transforms market data into 4-week price forecasts with full explainability.
Data Ingestion
Automated collection from 6+ primary sources every Tuesday. Raw data is validated, cleaned, and normalized.
Feature Engineering
48 price-driving signals are computed including lag features, rolling averages, and cross-source correlations.
Model Inference
LightGBM ensemble generates 4-week forecasts with confidence intervals for each prediction horizon.
Explainability
SHAP values computed for every prediction, showing exactly which factors drove the forecast up or down.
Weekly Update Schedule
Fresh forecasts published every Tuesday after Fish Pool spot price release
Machine Learning Model
LightGBM: Speed and accuracy for time-series forecasting
We use LightGBM (Light Gradient Boosting Machine), a state-of-the-art gradient boosting framework optimized for structured tabular data. Unlike neural networks, LightGBM excels at capturing non-linear relationships in small-to-medium datasets typical of commodity markets.
Feature Engineering
48 price-driving signals including lag features, rolling statistics, and cross-correlations
Training Regime
Weekly retraining on expanding window of historical data with walk-forward validation
Forecast Horizon
4 rolling weeks with separate models optimized for each horizon
Regularization
L1/L2 penalties and early stopping to prevent overfitting to recent noise
FEATURE IMPORTANCE (GAIN)
Feature importance calculated via split gain across all trees. Supply-side factors dominate short-term price movements.
SHAP WATERFALL - WEEK 14 FORECAST
Explainability
Every prediction tells a story
We believe forecasts should be transparent. Using SHAP (SHapley Additive exPlanations), we decompose every prediction into its component drivers, showing exactly why the model expects prices to move.
SHAP values come from game theory and provide mathematically consistent feature attributions. Green bars push prices up, red bars push them down. The sum of all contributions equals the final forecast.
Why SHAP matters for decision-making
Data Sources
Primary data from authoritative sources
We aggregate data from official government registries, market exchanges, and proprietary alternative data feeds to build a comprehensive market view.
Fiskeridirektoratet
GovernmentBiomass data, lice counts, and production reports from all 13 Norwegian production zones.
BarentsWatch
GovernmentReal-time aquaculture intelligence including sea temperatures, mortality data, and facility locations.
Fish Pool ASA
Market DataOfficial spot price index (NASDAQ Salmon Index) and forward curve data from the salmon derivatives market.
Statistics Norway (SSB)
GovernmentWeekly export volumes, trade flows, and historical price series for Norwegian salmon.
FRED (Federal Reserve)
EconomicMacroeconomic indicators including EUR/NOK exchange rates, commodity indices, and global economic data.
News Sentiment
Alternative DataAI-classified sentiment from industry news sources, detecting market-moving events and supply disruptions.
13 Norwegian Production Zones Monitored
From Rogaland in the south to Troms and Finnmark in the north, we track biomass levels, sea temperatures, and regulatory conditions across all official production areas.
Accuracy & Validation
Rigorous backtesting against real market data
We continuously validate our forecasts against the NASDAQ Salmon Index (SISALMONI) - the official benchmark for salmon spot prices used by Fish Pool ASA and the broader industry.
Our walk-forward validation methodology ensures that backtested performance reflects realistic out-of-sample conditions. We never peek at future data when evaluating historical accuracy.
87.5%
Direction Accuracy
Correctly predicted up/down
4.2%
MAPE
Mean absolute percentage error
SISALMONI
Benchmark
Official Fish Pool index
2 Years
Test Period
Walk-forward validation
10 of 12 direction calls correct
Model successfully predicted whether price would rise or fall
Explore Salmocast products
Price Forecasting
4-week salmon price forecasts with 87.5% directional accuracy. ML-powered predictions with confidence intervals and SHAP explainability.
OSINT Intelligence
Real-time production zone monitoring across all 13 Norwegian regions. Track biomass, lice levels, and supply pressure indicators.
Ready to see the forecast in action?
Experience transparent, explainable salmon price forecasting powered by machine learning and 48 price-driving signals.