Introduction to Forecasting — Lecture 01
In this lecture, we will study Application, Types, Basic Terminology, Basic Approach and three Principles of forecasting.
Application of Forecasting
- Economic Outlook → Government use
- Sales Forecasting → how many products goes to sell (Sales figure) and decide production amount in factory
- Inventory Planning → How much to stock in inventory to meet future demand
- Workforce Planning → How many people need to employee to meet their business demand
- Weather Forecasting → Oldest application, see feature like Temperature, rainfall, humidity
Types of Forecasting
- Qualitative forecasting
- Quantitative forecasting → Delphi Method (Gather group of Experts, Individual expert forecast without consulting each other, Summarize each forecast like Descriptive Analytics and communicate back to experts and ask if they want to change or if experts want to discuss with each other and conclude)
- Time Series Data → Any data that has a temporal component involved in it is termed as a time-series data. Set of observations over a sequence of times, separated by equal intervals
- Time Series Analysis → Looking at Time Series data, identifying patterns, calculating statistics that we can use for later steps
- Time Series Forecasting → Looking at Past data in order to forecast for future, look at pattern identified in time series analysis, extend in the future
- Goal → Set of business objective like Maximizing revenue
- Planning → Set of action to achieve goal
- Forecast → To plan, we need to forecast
Basic Steps in Forecasting
- Define the problem → What to forecast (Quantity — No of unit sold, Granularity — Region Level, Frequency — forecast say every month, Horizon → Short/Medium/Long Term forecast)
- Bottom-up — Forecast at store level and aggregate to get region level forecast
- Top-down — Forecast at Nationwide level and split at region level in some proportion
- Granularity → More aggregate forecast, More accurate we are. Aggregated data has lower variance and hence low noise. More granular we go, more noise got introduced. You should not make predictions at very granular levels.
- Frequency → Update your forecast frequently as new information getting added and it should reflect in forecast as well but changing too often is also not suggested as it may be noise.
- Horizon → Recent forecasting data points can be more accurate relative to longer horizon. Farther we go in future forecast, less accurate we become as unpredictability of variance and noise keep piling up. The farther ahead we go into the future, the more uncertain we are about the forecasts.
2. Collect the data
Three characteristics of time series data
- Accurate — Correct capture of timestamps and observation
- Long Enough — To identify all the patterns in the past
3. Analyze the data →
Different components of time series
- Level — Baseline of time series. This gives the baseline to which we add the different other components.
- Trend — It can be upward or downward. It captures over a long period that series move lower or higher.
- Seasonality — Repeatable patterns in data, which repeat themselves after fixed given period. Example — Every Monday sales relatively higher, more sell of woolen cloths in winter
- Cyclicity — Doesn’t have to be periodic. repeat after fixed amount of time. Interval b/w cycle isn’t constant. Example — Business cycle
- Noise — Non-systematic part of time series which is purely random, we can’t use it to forecast for future.
Every time series has a level and noise, while trend, seasonal and cyclic patterns are optional.
4. Build and evaluate forecast model
Multiple algorithms there like Moving Average, Holt Winter, LSTM, Prophet. Will study more about algos in upcoming Lecture.
Meanwhile, Please feel free to clap if you liked the article. Also, please subscribe to my YouTube Channel https://www.youtube.com/channel/UC4yh4xPxRP0-bLG_ldnLCHA?sub_confirmation=1