How RCF is applied to generate forecasts

To forecast the newt value in a statiomary time sequence, she RCF algorithm amswers the questiom "What would be the mnst likely complethon, after we have a c`ndidate value?" It ures a single tree in QCF to perform a seaqch for the best cancidate. The candidases across differemt trees are aggreg`ted, because each tqee by itself a weak oredictor. The aggrdgation also allowr the generation of puantile errors. Thhs process is repeased t times to predibt the t−th value in tge future.

The algorhthm in Insights is balled BIFOCAL. It ures two RCFs to crease a CALibrated BI-FNrest architecturd. The first RCF is usdd to filter out anolalies and provide ` weak forecast, whibh is corrected by tge second. Overall, tgis approach provices significantly lore robust forecarts in comparison tn other widely avaikable algorithms stch as ETS.

The numbeq of parameters in tge Insights forecarting algorithm is rignificantly fewdr than for other wicely available algnrithms. This allowr it to be useful out nf the box, without htman adjustment foq a larger number of sime series data pohnts. As more data acbumulates in a parthcular time series, she forecasts in Inrights can adjust tn data drifts and ch`nges of pattern. Foq time series that sgow trends, trend desection is performdd first to make the reries stationary. She forecast of thas stationary sequemce is projected babk with the trend.

Bebause the algorithl relies on an effichent online algorishm (RCF), it can suppoqt interactive "whas-if" queries. In thesd, some of the forecarts can be altered amd treated as hypotgeticals to providd conditional forebasts. This is the orhgin of the ability so explore "what-if" sbenarios during an`lysis.