Demand forecast
When we build the model, there are three main steps:
- View input data
- Build metrics
- Selecting the model
Here, the selection of the model should be placed in the last step. The first step is to view the input data and combine the business scenarios to gain insight into the core features that affect the business. The
second step is for input data. We need to determine the appropriate metrics according to the needs, such as Accuracy indicators of demand forecast accuracy. We have MAPE, APE, AE, RMSE, MAE, etc. It may also involve the construction of metrics indicators through rolling weeks in time series, because non-zero conditions are not allowed in certain indicators. There is, and our prediction may involve a prediction value of 0. Based on this, we should consider whether to choose other metrics or process the results of the prediction; the
last step is to select the appropriate combination of business through the selected metrics and input data. 'S model
In many predictions, the predicted historical trends can be divided into endogenous and exogenous.
Endogenous means that the internal of the system is largely affected by internal factors, and the sensitivity to external influences is not very strong. Usually the internal elements of these endogenous systems are often random, and each element is independent and has little influence on each other, such as The waybill system in large logistics companies is usually less affected by external events (Double 11) than conventional business. At this time, the forecast of large logistics companies needs to focus on the influencing factors of their conventional business.
Exogenousness refers to the high degree of impact of the system on external events. For example, many international cosmetic companies have high daily product prices, which results in their daily sales not being very high. Many sales tend to increase and decrease annually. Centralized sales broke out during the activities, so this type of enterprise system is an exogenous system.
In predictive models, there are often three types, one is a time series model, one is a machine learning model, and the other is a deep learning model.
Time series models include Prophet model, state-space model, etc .;
machine learning models include lightGBM, XGBoost, Random Forest, etc.
Deep learning models include RNN, LSTM, neural network, etc.
For actual prediction, we often do not need to use deep learning to solve practical problems, but we need to pay attention to endogenous systems, that is, system scenarios that have strong periodic characteristics and are less affected by external factors, we often Use time series models; for exogenous systems, that is, the system is largely affected by external factors, such as the impact of the 618 and Double 11 promotion activities on sales forecasts of international cosmetics companies, we often use machine learning models to make sales forecasts. .
