Revise definition of "historical variability" in direction/trend calculations

[krivard]

The current definition of trend/direction computes a linear regression of the past N days (N=7, N=14) and then classifies the curve as decreasing, steady, or increasing by comparing the slope of the line to a measure of historical variability computed using the data available for that signal before 22 April. Since not all signals existed before 22 April, this prevents us from publishing direction/trend on some signals.

Wael identified this issue during testing for the quick performance fixes;

Some ideas for fixes, none of them good, for deciding historical variability:

• When historical data before 22 April is unavailable, default to some constant (not statistically feasible, since signals are on wildly different scales)
• When historical data before 22 April is unavailable, compute using the first month of available data, whatever that is (not computationally feasible with the amount of data we have now)
• Compute historical variability for all signals using the last 30 days from the day the direction updater is run (not stable over time)

Additional past discussion

[krivard]

We're moving direction computation out of the backend and into the client.