1
$\begingroup$

I have a set of data, let's say average weight of employees, captured every month over a period of 5 years (2010 - 2014). I cannot find a seasonality trend in the data over these years. Also, I have found that it is not dependent on any other factors.

I am trying to forecast values for 2015 to get a general sense of this data as it is an important metric in the operations of my business.

I have tried ARIMA, R-regression, Exponential smoothing, Excel forecast to find any seasonality whatsoever. However, my efforts are yet to materialize.

My question is: How do I forecast a variable that has no seasonality?

I have attached my data herewith.

Graphs

Yearly Values for years 2010 - 2014

enter image description here

Value Cumulative over 2010 - 2014

enter image description here

All Values from 2010 - 2014

enter image description here

Auto ARIMA in R

# Map 1-based optional input ports to variables
dataset1 <- maml.mapInputPort(1) # class: data.frame
library(forecast)


dates <-  dataset1$Date
values <- dataset1$Weight

dates <-  as.Date(dates, format = '%m/%d/%Y')
values <- as.numeric(values)

train_ts <- ts(values, frequency=12)
fit1 <- auto.arima(train_ts)
train_model <- forecast(fit1, h = 12)
plot(train_model)

# produce forecasting
train_pred <- round(train_model$mean,2)
data.forecast <- as.data.frame(t(train_pred))
#colnames(dataset1.forecast) <- paste("Forecast", 1:data$horizon, sep="")

# Select data.frame to be sent to the output Dataset port
maml.mapOutputPort("data.forecast");

Forecasted Value with Auto ARIMA

Date        Weight
01-01-15    11.77
01-02-15    11.76
01-03-15    11.77
01-04-15    11.76
01-05-15    11.77
01-06-15    11.77
01-07-15    11.76
01-08-15    11.77
01-09-15    11.76
01-10-15    11.77
01-11-15    11.77
01-12-15    11.76

Data

Date        Weight      Cumulative Weight
01-01-10    11.8800     11.8800
01-02-10    10.4000     22.2800
01-03-10    6.9500      29.2300
01-04-10    15.5000     44.7300
01-05-10    17.0400     61.7700
01-06-10    10.4700     72.2400
01-07-10    12.1400     84.3800
01-08-10    2.5800      86.9600
01-09-10    12.6300     99.5900
01-10-10    11.6800     111.2700
01-11-10    9.0700      120.3400
01-12-10    10.8900     131.2300
01-01-11    1.7500      132.9800
01-02-11    -1.7700     131.2100
01-03-11    5.9300      137.1400
01-04-11    -4.9200     132.2200
01-05-11    4.3900      136.6100
01-06-11    1.5100      138.1200
01-07-11    1.2200      139.3400
01-08-11    10.2900     149.6300
01-09-11    13.0600     162.6900
01-10-11    10.1400     172.8300
01-11-11    8.5250      181.3550
01-12-11    6.4350      187.7900
01-01-12    -5.5100     182.2800
01-02-12    -4.3000     177.9800
01-03-12    2.3200      180.3000
01-04-12    4.0700      184.3700
01-05-12    12.2700     196.6400
01-06-12    14.7400     211.3800
01-07-12    8.4600      219.8400
01-08-12    11.6300     231.4700
01-09-12    -0.1500     231.3200
01-10-12    2.5200      233.8400
01-11-12    6.7400      240.5800
01-12-12    35.6300     276.2100
01-01-13    26.4000     302.6100
01-02-13    26.1300     328.7400
01-03-13    16.2100     344.9500
01-04-13    56.0800     401.0300
01-05-13    32.2300     433.2600
01-06-13    17.5100     450.7700
01-07-13    3.6700      454.4400
01-08-13    7.7700      462.2100
01-09-13    -14.2800    447.9300
01-10-13    1.0800      449.0100
01-11-13    9.4000      458.4100
01-12-13    7.3400      465.7500
01-01-14    6.1400      471.8900
01-02-14    3.8200      475.7100
01-03-14    16.7600     492.4700
01-04-14    0.4900      492.9600
01-05-14    17.9800     510.9400
01-06-14    14.8000     525.7400
01-07-14    12.6400     538.3800
01-08-14    5.7300      544.1100
01-09-14    -2.0900     542.0200
01-10-14    9.1300      551.1500
01-11-14    12.5100     563.6600
01-12-14    -1.3900     562.2700

Actual Values for 2015

Date        Weight
01-01-15    -18.43
01-02-15    13.94
01-03-15    26.14
01-04-15    24.36
01-05-15    18.37
$\endgroup$
  • $\begingroup$ I am a bit confused, have you answered your question? $\endgroup$ – TPArrow Jul 1 '15 at 9:33
  • $\begingroup$ @Hamed I have tried with ARIMA and that's the best result I have got. I wanted to know if there is anything better? $\endgroup$ – AngryPanda Jul 1 '15 at 9:35
  • $\begingroup$ fine, I guess the most common method is ARIMA which you have already used. Lets wait fr more answers $\endgroup$ – TPArrow Jul 1 '15 at 9:37
2
$\begingroup$

It does make sense that there is no seasonality in weight data. Yes, people may eat more over Christmas and try to slim down for the beach, but these are really very minor effects. (And if you don't actually have weight data, but other series, these still can be nonseasonal. Like sales of toilet paper: the, ehm, data-generating process runs year-round with little variation.)

If you don't have seasonality, what else can there be? There could be cycles of varying and unstable lengths, like economic cycles. There could be trends. There could be changes in variance. (You already wrote that there does not seem to be a causal relationship with other variables, so I'll exclude causal effects.)

If you have a trend, like with your cumulative weight data, auto.arima() should detect it and apply differencing to arrive at stationary data. (In effect, it would be undoing the cumulation to be back at the stationary separate weight data points.)

In any case, after differencing, you are left with (we hope) stationary and non-seasonal data. Then auto.arima() will fit an appropriate ARMA model. A weight time series will likely be autoregressive, so you may end up with an AR(1) or AR(2) or similar model. The point forecasts from such models will depend on previous values in an autoregressive way, and they quite likely won't change a lot.

It's also quite possible that auto.arima() yields a zero-th order model, i.e., a simple mean model, with forecasts that are a flat horizontal line - and this may indeed be the best forecast you can make, if there is simply no ARMA structure to be modeled.

Bottom line: if your series have no obvious seasonality, then your forecasts will quite likely not be seasonal, either (why should they?), and can reasonably be an essentially or totally flat line.

I like recommending this free online forecasting textbook.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.