Octopus Watch can provide electricity rate predictions for Agile users. These predictions allow you to plan ahead, even when official prices aren’t announced yet. Predictions are shown as a dotted line ⋯ on the rates chart. Predictions are continuously updated to show the most recent and most accurate predictions.
Predictions are an advanced feature that is only available to subscription users.
How to use Predictions
The chart with the dotted line can be used as-is for a quick indicator of future prices. Yet a little knowledge goes a long way to better use these predictions to benefit you:
- 3–7 days ahead
- Use the weather chart which you can enable in Settings → Options. The weather chart shows you wind (blue) and temperature (orange) across Great-Britain (GB). The thin wobbly line is the average, and the shaded area around it is a confidence band – the highs and lows across GB. There are two horizontal barriers marked with a thumbs up and thumbs down. If the blue line (and its confidence band) goes above the blue horizontal line then Agile prices tend to be very low because there is a lot of wind. If the orange line drops below the orange horizontal line then Agile prices tend to be higher as people are using lots of energy to heat their properties. Starting in version 4.3.4: if the orange line goes over the blue line prices also tend to be worse as lots of people turn to air-conditioning.
- from 48h in advance
- The half hourly predictions come into play. The most important predictor used by this model is still the weather, though other information is starting to trickle in too. Like a UK weather forecast these price predictions can shift around quite a bit. Use these predictions as an indication, but don’t use them to plan ahead. The default in the app is to only show 24h predictions. The slot finder, even if you enable the option to use predictions, will never use these 48h predictions to find a slot.
- from 24h in advance
- As more data comes in better models take over to offer more accurate predictions. These rate predictions are often reliable. If you want, you can enable the option to use these predictions to find the cheapest slot.
- the day before, from 9-11AM
- More and more balancing information for the National Grid becomes available. This allows very accurate predictions that are often close to the actual values. Most of the error is akin to being a bit higher for one half hour and lower for another.
- the day before, from 4PM
- All balancing data is available and prices can be estimated highly accurately. Next day values are typically loaded into the app by 3:55PM. The app, thanks to its predictions, can show you next day rates even if the rates do not become available through Octopus Energy at 4PM.
Price plunges – where prices drop below 0p/kWh – are unexpected events, such as a sudden change in wind speed. These cannot be predicted.
Nix, the prediction engine, is currently in its 3rd iteration. This sophisticated tool has been refined over the years and uses a huge collection of data to predict upcoming prices. There are a few key properties to understand how it works.
Nix is a multi-model engine :it doesn’t just train one model, but it is training multiple models. Currently, Nix uses 4 distinct models corresponding with a 48h model, a 24h model, an 8h model, and a 0h model. Each model has access to more data and uses this extra data to create more accurate predictions.
Nix continuously retrains :which means that it is always learning from the latest data. This allows it to adapt to changes in the wholesale market such as in 2022. Before a new model is adapted it does have to out-compete the old model. This is to ensure that the new model is actually better than the old one.
Nix predicts wholesale electricity prices :and not just Agile prices. This gives Nix great flexibility to adjust what comes our way – such as sudden discounts like the Energy Price Guarantee.
Nix has a mood :This mood influences its predictions and is used to make the 24h model more pessimistic. As a result, prices are typically predicted higher, thus minimising potential disappointment should the actual prices turn out to be higher than expected.
Nix uses a lot of data :Approximately 450 features/columns and ~125,000 points/rows are used for training. Some examples of features are e.g. the wind speed across GB, the humidity in London, the expecte nuclear output, or whether it is a bank holiday.
How accurate are the predictions?
At the end of the day what matters most is the accuracy of the predictions. The performance of Nix is continuously tracked and evaluated. Averages lie and most metrics in statistics and artificial intelligence are best used by experts because they only draw a part of the picture. Instead, below you can find two randomly selected dates where the predictions are overlayed on top of the actual prices. There are two caveats: the wholesale prices are reshaped to match Agile prices on the given dates, and Nix used the actual weather data instead of the weather data forecast. This mostly affects model 1, which would perform worse (i.e. have bigger errors) when the forecast itself is off.