Insights Report - Forecasting
Last updated
Last updated
The insights report fields vary based on the data input and the model type. This page covers an example of the forecasting model and what data it outputs. The dataset used here is an overview of energy demand in NYC over time.
A breakdown of your overall results from training the forecasting model. This shows the overall prediction accuracy of the control data ( in the case of Forecasting, the most recent data is reserved) and the confidence bounds of the set.
Select 'See accuracy details' to expand the Predictive Performance deep dive. You can see the prediction line overlayed on the training data. In the case of the example dataset, the data curve is clearly followed, and higher peaks are predicted more accurately. There are three values calculated for the data:
Accuracy - Predictions are usually within this percentage (plus or minus) of the actual outcome. Lower is better.
Root Mean Square Error - measures how correct the predictions are on average. It is calculated by measuring how far away the predicted values are from the true values. Lower is better.
Mean Absolute Error - is a common measure of forecast error in time series analysis. It is the mean absolute value of the difference between predictions and actuals. This helps you understand the average size of prediction errors without considering if they are above or below actuals. Lower is better.
Accuracy Decay shows the decrease in certainty of the model over time. Any time-based forecast will decrease in accuracy as you get further forward. Retraining the model as current data comes in ensures the model stays accurate going forward.
A graphical representation of the training data vs the predicted data. You can optionally show the confidence interval to see how strong the predictions are at any particular point.
To receive a spreadsheet of the predicted values, select "Download Dataset" on the Forecast.
Seasonality allows the user to break the data down into specific time windows to see how the value of interest(in this case, demand) changes based on the time of day, day of the week, month of the year, quarter, etc.
Akkio leverages two key approaches to analyze time-series data: impact analysis and lagged correlation calculation.
Impact analysis allows us to model specific scenarios by quantifying the expected change after a certain lag period in a dependent variable, given a specific change in an independent variable. This approach provides insights into causal dynamics, which can be critical in understanding complex systems. The impact analysis aims to highlight specific and meaningful relationships that might be non-linear (and might not be surfaced by a regular lagged correlation analysis).
To view correlations click the bar chart icon in the upper right of the Leading Indicators section. Lagged correlations measure the statistical relationship between two variables over time. This is done by aggregating the data over that period by computing the correlation between two time-series datasets at different points in time after downsampling to the appropriate time period.
