How to measure the growth of a trend in time-series analysis?

Question

I have a time series dataset as follows

day                1     2      3     4      5     6

sales_dept1        0     0      0     3      3     3
sales_dept2        0     0      0     6      6     6
sales_dept3        0     0      0     1      1     1

As you can see, dept2 is most efficient, dept1 is second most efficient and dept3 is less efficient.

In other words, I want to capture the performance diffrences among the departments.

For, that I used Linear Regression and got the slope as a metric to compare the performace. However, I got the following results as the slope (which do not match with my manual interpretation).

dept1 slope = 1
dept2 slope = 0.5
dept3 slope = 3

My code is as follows.

from sklearn.linear_model import LinearRegression
regressor = LinearRegression()

X = [[0], [0], [0], [3], [3], [3]]
y = [1, 2, 3, 4, 5, 6]
regressor.fit(X, y)
print(regressor.coef_)

Therefore, I think slope is not the correct metric for my situation. Is there any other metric that captures the actual growth change of different time-series data?

I am happy to provide more details if needed.

Update:

Mentioned below are some more data:

[0, 0, 0, 1, 1, 1]
[6, 6, 6, 6, 6, 6]
[0, 0, 0, 0, 0, 10]
[0, 3, 3, 28, 30, 30]
[0, 0, 0, 6, 6, 6]
[0, 0, 0, 0, 0, 10]
[0, 0, 0, 0, 1, 1]
[6, 6, 6, 6, 6, 6]
[0, 0, 0, 0, 1, 1]
[0, 0, 0, 6, 6, 6]
[0, 1, 1, 4, 4, 4]
[3, 19, 19, 47, 64, 90]
[0, 0, 3, 8, 13, 13]
[0, 3, 3, 3, 3, 3]
[0, 0, 0, 6, 6, 6]
[0, 0, 0, 0, 0, 6]
[0, 0, 0, 0, 0, 0]
[0, 0, 0, 0, 0, 10]
[0, 0, 0, 0, 0, 6]
[10, 10, 10, 10, 10, 10]
[0, 0, 0, 0, 10, 10]
[0, 0, 0, 7, 7, 15]
[0, 0, 0, 0, 10, 10]
[6, 6, 6, 6, 9, 9]
[6, 6, 6, 6, 6, 6]
[0, 0, 3, 3, 3, 3]
[0, 0, 0, 6, 7, 7]
[0, 3, 4, 4, 4, 4]
[0, 0, 0, 0, 1, 1]
[3, 3, 3, 3, 3, 9]
[0, 0, 6, 6, 6, 6]
[0, 0, 0, 0, 0, 0]
[0, 0, 0, 6, 6, 6]
[1, 1, 1, 1, 1, 1]
[0, 0, 0, 15, 15, 15]
[0, 15, 15, 15, 15, 15]
[0, 0, 0, 0, 0, 0]
[0, 1, 1, 1, 1, 1]
[0, 0, 1, 1, 1, 1]
[0, 15, 16, 22, 32, 68]
[0, 0, 0, 1, 1, 1]
[0, 0, 0, 0, 1, 1]
[0, 3, 3, 4, 4, 4]
[0, 0, 0, 15, 15, 15]
[0, 0, 0, 0, 0, 10]
[0, 0, 0, 21, 21, 21]
[0, 0, 0, 6, 6, 6]
[0, 0, 0, 9, 9, 24]
[0, 0, 0, 3, 3, 3]
[0, 0, 0, 0, 6, 6]
[0, 3, 3, 3, 3, 3]
[0, 0, 11, 11, 19, 19]
[3, 18, 18, 43, 50, 76]
[0, 0, 0, 0, 0, 15]
[0, 0, 0, 0, 0, 0]
[14, 14, 14, 39, 42, 72]
[6, 6, 6, 6, 6, 6]
[0, 0, 0, 0, 3, 3]
[0, 0, 0, 0, 0, 3]
[0, 0, 0, 0, 6, 6]
[0, 0, 0, 15, 15, 15]
[0, 0, 6, 6, 6, 6]
[0, 0, 10, 10, 10, 10]
[0, 0, 6, 6, 6, 6]
[0, 0, 10, 10, 13, 13]
[0, 0, 0, 6, 11, 11]
[6, 6, 6, 6, 6, 6]
[0, 0, 0, 0, 3, 3]
[0, 0, 0, 0, 0, 0]
[10, 10, 10, 10, 10, 28]
[0, 0, 0, 6, 6, 6]
[0, 0, 0, 15, 15, 15]
[0, 0, 0, 0, 1, 1]
[0, 0, 0, 0, 0, 10]
[1, 1, 1, 1, 1, 1]
[0, 0, 0, 6, 6, 6]
[0, 0, 0, 21, 21, 21]
[0, 0, 0, 0, 0, 15]
[0, 0, 0, 0, 0, 6]

You want to capture performance differences among the departments. What do you mean by good performance? Faster growth than others? (slope of a deterministic linear trend measures it, but is the trend linear?) Or higher average sales? Anyway six day is really few to speak about any kind of trend. — oszkar, Mar 20 '19 at 10:57
@oszkar thanks for the comment. we do weekly analysis. that is why I only have 6 days :( do you have any suggestions? yes, I want to identify what departments are faster growing than other departments. Looking forward to hearing from you :) — EmJ, Mar 20 '19 at 12:22
@oszkar It is not simply the # of observations but rather the ratio of signal to noise. A 6 valued series 1,2,3,4,5,6 speaks to a trend . 1,1,1,3,3,3 speaks to a change in level/step ... 1,2,1,2,1,2 speaks to a pattern ... 1,1,1,1,1,2 speaks to a pulse,... As usual of these examples/conclusions may be "proven" to be right or wrong as future values become known. — IrishStat, Mar 20 '19 at 17:25
@EMI If you have more data please modify your post to include all data to date. — IrishStat, Mar 26 '19 at 12:53

IrishStat · Accepted Answer · 2019-03-20T15:53:57.967

.Trends can be of two forms y(t)=y(t−1)+θ0 (A) Stochastic Trend or Y(t)=a+bx1+cx2 (B) Deterministic Trend etc where x1=1,2,3,4....t and x2=0,0,0,0,0,1,2,3,4 thus one trend applies to observations 1−t and a second trend applies to observations 6 to t. Often software is needed to aid the detection of the trend changes.

If you post a real series I may be able to help you further.

Trend shifts in timeseries provides some insight/guidance .

Apparently you are tracking some 79 items over time ( currently 6 days )

In the future please attach an external csv file to your post reflecting the total history of the 79 items in the form presented here for ease of analysis.

For example I took series 11 and obtained

EDITED AFTER OP'S QUESTION AS TO HOW THIS GRAPH WAS DRAWN:

After considering two options :

1) Build a model using a PREDICTOR SERIES of the form 1,2,3,4,5,6

2) Build a model optimally using the history of prior values (ARIMA)

the tournament ( i.e. set of trials) concluded the the best model was 2) which meant that the prediction at any one point would be to use the most recent value and add 17.4 . The graph shows the Actual and the 1 period out prediction at each point in time IN GREEN.

I have taken your 79 series of length 6 and created 79 png files showing the actual and the fitted/predicted values which represent the equation . If you contact me at dave@autobox.com , I will be happy to send them to you along with the companion equation files. I just don't know how to attach them to this post.

Thanks a lot. I only have the data for 6 days. I will post more data now. Looking forward to hearing from you. :) — EmJ, Mar 20 '19 at 06:55
I updated the question. Please let me know your thoughts. Thank you very much once again :) — EmJ, Mar 20 '19 at 07:00
thanks a lot for the update. However, I am still not clear how this graph is drawn. Can you please tell me some details about the method you used? :) — EmJ, Mar 20 '19 at 12:20

oszkar · Answer 2 · 2019-09-03T08:02:49.270

1

I would do the next for a fast and simple solution:

Iterate through the departments, calculate the slope of the linear regression.
Create groups/clusters from the slopes to have performance groups. You can apply some cluster analysis, some initially set thresholds, or just a check the results and use common sense. For some automated solution the first two approach could be use.

(Later will add some example code to show what I mean.)

edited Sep 03 '19 at 08:02

answered Mar 20 '19 at 13:17

oszkar

736
4
15

thank you very much. However, as I have mnetioned in the question regression slope did not give me the output I expected. Looking forward to hearing from you :) yes, it would be really helpful if you could add some example code. Thank you very much once again :) – EmJ Mar 20 '19 at 13:21
1

@Emi Will check that too. – oszkar Mar 20 '19 at 13:40
thanks a lot. i look forward to hearing from you :) – EmJ Mar 20 '19 at 15:54
If there are multiple trend changes or level/step changes or pulses or autoregressive structure in the data your suggestion of using the slope of a simple linear regression on time is way to simple and ignores the work of time series analysts for the last 60 years – IrishStat Mar 22 '19 at 00:38
@IrishStat As far as I get it right no trend changes was in question: "yes, I want to identify what departments are faster growing than other departments". (The "growth of a trend" part in the title could be misleading.) As far as I understand, it want to be measured within a week. But Emi will clear that. – oszkar Mar 26 '19 at 08:41

How to measure the growth of a trend in time-series analysis?

2 Answers2