在 Oracle/SQL 中预测时间序列数据

Question

Is there a way we can generate a time series forecasting for a data set using an Oracle analytical functions? How do we perform extrapolation in SQL/ORACLE.

Below is my need

I have data data set like below and I wanted to forecast/extrapolate for next year

Cust_id  Year  Revnue
1        2016  679862
1        2017  705365
1        2018  ?
2        2016  51074
2        2017  50611
2        2018  ?
3        2016  190706
3        2017  90393
3        2018  ?
4        2016  31649
4        2017  19566
4        2018  ?

Answer 1

You can create a simple forecast using the REGR linear regression functions.

--Ordinary least squares forecast for each customer for the next year.
select
    cust_id,
    max(year) +1 forecast_year,
    -- y = mx+b
    regr_slope(revenue, year)
        * (max(year) + 1)
        + regr_intercept(revenue, year) forecasted_revenue
from customer_data
group by cust_id;

CUST_ID   FORECAST_YEAR   FORECASTED_REVENUE
-------   -------------   ------------------
1                  2018               730868
2                  2018                50148
4                  2018                 7483
3                  2018                -9920

Below is the sample schema. Or you can use this SQLFiddle .

create table customer_data
(
    cust_id number,
    year number,
    revenue number
);

insert into customer_data
select 1, 2016, 679862 from dual union all
select 1, 2017, 705365 from dual union all
select 2, 2016, 51074  from dual union all
select 2, 2017, 50611  from dual union all
select 3, 2016, 190706 from dual union all
select 3, 2017, 90393  from dual union all
select 4, 2016, 31649  from dual union all
select 4, 2017, 19566  from dual;

---

The REGR function deals with number pairs, it doesn't understand business rules like "revenue can't be below 0". If you want to restrict the forecasts to always stay at or above 0, a CASE expression may help:

--Forecasted revenue, with minimum forecast of 0.
select cust_id, forecast_year,
    case when forecasted_revenue < 0 then 0 else forecasted_revenue end forecasted_revenue
from
(
    --Ordinary least squares forecast for each customer for the next year.
    select
        cust_id,
        max(year) +1 forecast_year,
        -- y = mx+b
        regr_slope(revenue, year)
            * (max(year) + 1)
            + regr_intercept(revenue, year) forecasted_revenue
    from customer_data
    group by cust_id
);

CUST_ID   FORECAST_YEAR   FORECASTED_REVENUE
-------   -------------   ------------------
1                  2018               730868
2                  2018                50148
4                  2018                 7483
3                  2018                    0