Computer-assisted decision-making
Oljefondet analyzes large amounts of data continuously and delivers weighted recommendations to investors and analysts who want to reduce risk on a verifiable basis, not on gut feeling.
Methodology
Most decision-making tools stop at the visualization of data. Oljefondet goes further and assigns each recommendation a weight based on how reliable the signal has historically proven to be.
The models are continuously trained on historical and updated data sets, and discover relationships between variables that would otherwise require manual analysis over weeks. The result is presented as probability-weighted scenarios, not as a single recommendation without context.
Illustration: weighted probability per scenario
Each recommendation is followed by a correlation check against existing exposure and a confidence value that shows how much historical deviation has been observed for similar signals. This makes it possible to prioritize measures according to actual downside risk, rather than magnitude alone.
Illustration: deviation distribution per risk category
Because the architecture is built on flowing data, recommendations are updated as new data points come in. The same infrastructure is used regardless of whether it is one analyst or a whole team monitoring several positions at the same time.
Illustration: update frequency over time
Technical rigor
The process is divided into three steps, and each step is documented so that analysts can verify why a recommendation looks the way it does.
01
Market data, macroeconomic indicators and alternative data sets are collected continuously and cleaned of known sources of error before proceeding to modelling.
02
Statistical and machine-learned models look for recurring relationships and deviations from normal distribution, and flag signals that cross a predefined threshold.
03
Signals are ranked according to historical accuracy and compiled into a recommendation with an associated confidence value, ready for human assessment.
| Parameter | Description |
|---|---|
| Data sources | Market data, macroeconomic indicators, publicly available company reports |
| Update frequency | Continuous, based on streaming data rather than periodic batches |
| Model architecture | Ensembles of statistical and machine-learned models with separate validation |
| Validation method | Verification against actual outcome, published continuously in the performance log |
| Integration | API access and export to common analysis and portfolio tools |
About the platform
Oljefondet was developed with a simple starting point: a recommendation is only as good as the ability to check it afterwards. Therefore, each model is documented, and each recommendation is time-stamped.
The platform is used by analysts and investors who already have their own routines for risk management, and who want a computer-assisted supplement rather than a replacement for professional assessment.
Transparency
All recommendations are recorded with time and later compared with the actual outcome. The log is available for review and cannot be edited after publication.
The table below shows the format of the log. A complete and continuously updated log is available inside the platform.
| Date | Model | Category | Status |
|---|---|---|---|
| 2024-01-08 | Sentiment model A | Market sentiment | Confirmed |
| 2024-01-15 | Risk model C | Portfolio risk | Under assessment |
| 2024-01-22 | Trend model B | Sector rotation | Confirmed |
| 2024-01-29 | Sentiment model A | Market sentiment | Deviation registered |
Traceability
Per model
Editing after publication
Not allowed
Registration of outcomes
Regardless of model
Historical accuracy is shown in the platform as a line diagram broken down per model and time period, so that deviations between recommendation and actual outcome can be followed over time instead of being summarized in a single number.
Recommendations are time-stamped upon publication. The actual outcome is recorded afterwards and is automatically linked to the original recommendation, regardless of how the outcome turned out.
Areas of use
The same underlying models are adapted to different decision-making situations, depending on whether the purpose is portfolio management, market assessment or internal operations.
The models identify correlations between existing positions and propose adjustments that reduce overall exposure to individual events, without requiring a complete restructuring of the portfolio.
Large amounts of text and transaction data are compiled into a sentiment measure that is updated continuously, and which can be compared against historical turning points in the market.
Internal data on resource use and throughput is analyzed to uncover bottlenecks, so that decisions about prioritization can be made based on actual patterns rather than assumptions.
Questions and answers
Data loaded for analysis is stored separately from training data used to improve the models and is not shared between customers. Access control and storage location can be reviewed as part of a technical assessment before integration.
The models are continuously tested against new data that was not included in the training, and deviations from the expected accuracy are automatically flagged. Because the results are published in the performance log, systematic biases can be identified over time rather than remaining invisible.
Most setups start with exporting recommendations to existing analytics tools via API, without the need to change existing workflows. Deeper integration, such as direct connection to portfolio systems, is assessed separately per case.
Yes. The recommendations are intended as decision support, and each user can adjust the weighting or ignore individual signals. Such adjustments are logged separately from the model's original recommendation to preserve verifiability.
The outcome is recorded in the performance log in the same way as hits, and is used to adjust the weighting of the signal in question going forward. Failed recommendations are not removed from the log.
Start by going through the performance log and assess the accuracy yourself, before you connect your own data. Most teams start on a smaller scale and expand after comparing results over a few months.
