Variance Routing by Horizon

From WFM Labs

Variance routing by horizon is the method by which a planning system decides which of its loops should answer a gap between the capacity it has and the capacity it needs. The gap in any week is a sum of components with different causes and different persistence: a one-week spike, a seasonal week the plan already knew about, a level shift that has lasted a month, a drift in handle time, a run of attrition. Each has a lever whose delay fits it, and the method assigns each component to the slowest loop whose delay matches the component's persistence. Its one hard rule is that nothing persistent is allowed to be carried by the interval loop, because the interval loop can carry anything for a while and the slow loops never learn.

The page belongs to the series that begins at Workforce Planning as a Control System, where this method is coupling L3.

Why routing is needed

In a conventional cascade, variance is absorbed where it lands. A week eight percent over forecast is met with overtime and occupancy because those are the levers the week has. If the eight percent persists for six weeks, the same levers meet it six times, and the gap never reaches the horizon whose lever is hiring, because that horizon reconciles to the annual plan and the plan has not changed. The operation runs hot, the fast loop's cost accumulates as overtime and attrition, and the slow loop sees nothing it recognises as its own.

The fast loops are good at hiding this. A free real-time absorb, an overtime band, a deferrable share of work and some room to squeeze handle time can carry a sustained shortfall for many weeks at a cost that shows up as overtime and repeat contacts rather than as missed service; in the companion simulation the overtime and time-off band and the real-time absorb together carry about six-tenths of the weekly gap-hours under both the cascade and the closed loop [estimated]. That would explain why a cascade's open loops are rarely blamed for anything. Routing exists to make the slow loops' share of the variance explicit and to hand it to them.

The method

Four steps, run at the cadence of the short-term loop.

  1. Decompose the week's gap by cause. Level (contacts above or below the forecast), mix (the work arriving in a different pool than planned), handle time (the same contacts taking longer), shrinkage (fewer productive hours per paid hour than assumed), and supply (fewer people than the plan of record, from attrition, delayed starts or training loss). The decomposition is arithmetic on the capacity equation and is the same one the forecast-versus-actual review already performs; see Forecast Vintages for the version each comparison is made against.
  2. Measure each component's persistence. How many consecutive weekly cycles has it had the same sign? A component that has flipped sign inside the window is transient. One that has held its sign for the window, four consecutive weekly cycles in the companion simulation, is persistent at the capacity scale. One that has held for a pipeline length is persistent at the structural scale.
  3. Assign each component to a loop. Transient components stay with the interval and short-term loops. Components that have persisted for the window are handed to the capacity loop, where they enter the gap that the hiring rule closes, net of the supply line. Components persistent beyond one pipeline are handed to the structural loop, where they are a question about contract bands, cross-training and location mix rather than about headcount.
  4. Report what each loop carried. The share of the week's gap-hours carried by real-time actions, by the hours band, by the contracted band, by deferral, by handle-time compression, by unmet contacts and by hiring. The shares are the system's own diagnosis: a large share carried by handle-time compression is quality erosion in progress, and a large share carried by the hours band for a quarter is a capacity decision that has not been taken.

The persistence test is also what keeps the capacity loop from being nervous. A gap that moves the plan of record only after it has held its sign for the window, and then moves it by the smallest excursion in that window, filters the noise before it reaches the pipeline; the assumption ledger applies the same rule when a planned value is re-set. In the simulation the closed-loop policy releases to hiring only the part of the forecast gap that has persisted for four weeks; the price is a few weeks of lateness under a large step, and the return is about a quarter fewer hiring moves across a seasonal year (719 against 943), as one output of the combined closed-loop policy rather than of the persistence test alone [estimated]. See Planning Nervousness and Frozen Zones.

A worked decomposition

An illustration in indices, average week = 100, from one application of the method [estimated]. In a given week the operation needs 112 and has 100. The decomposition finds: level +7 (contacts above forecast for the fifth week), mix +2 (a pool over its forecast while another is under), handle time +3 (the standard being exceeded for the third week), supply −2 (a class that started late, so there are fewer agents than the plan). Persistence: level has held for five weekly cycles, which is past the four-week window and inside one pipeline; mix flipped sign last week; handle time has held for three cycles; supply is a known one-off.

Routing: mix and the one-off supply shortfall stay with the week, met by the hours band and cross-skill routing. The level component is handed to the capacity loop, which adds it to the thirteen-week gap net of the supply line, where it becomes a hiring decision, since it has already held for five weeks. The handle-time component has held for three and waits one more. When it passes the window the handle-time component is handed to the capacity loop too, but as an assumption breach rather than a hiring request: it goes to the assumption ledger, where the handle-time assumption's tolerance band has been exceeded, which is what opens a decision about whether the standard, the work or the training has changed; see Assumption Ledgers and the Continuous Annual Plan. Nothing is handed to the structural loop this week; if the level component is still present at the pipeline length, the question becomes whether the swing should be carried by a contract band rather than by hires.

What routing changes

The visible change is in the reports. The forecast-versus-actual review stops ending in an explanation and starts ending in an assignment: this component to this loop, by this rule. The capacity cycle's deficit stops being "plan minus floor" and starts being "the persisted components, net of the supply line". The interval loop's overtime report acquires a column for the variance it should not be carrying.

The structural change is that the slow loops begin to learn. A capacity loop that receives persisted variance by rule, rather than by escalation, sees the step change in the month it happens instead of the quarter it is reconciled. A structural loop that receives the variance still present at the pipeline length sees, for the first time, the swing that its contract bands were supposed to carry and did not.

Limitations

The decomposition requires one definition of a productive hour and one headcount, read from the same data by every loop; where those are missing the components are artifacts of arithmetic rather than of staffing, which is the "blended currency" failure in Capacity Planning Cycle. The persistence thresholds, a four-week window and one pipeline, are design choices chosen in the companion simulation; the nervousness literature supports filtering in general, not these thresholds.[1] The logic of closing a gap over an adjustment time rather than at once is the stock-management structure of system dynamics.[2] The method assigns variance to loops; it does not tell a loop what to do with it, and a capacity loop that receives a persisted gap and reconciles it to the annual plan has received nothing.

Maturity Model Position

Routing presupposes the decomposition, which is a Level 4: Planning in Distributions capability, and the loops it routes to, which are the control-system reading of the planning process. At lower levels of the WFM Labs Maturity Model™ the interval loop carries everything and the decomposition is done once a month by hand; at Level 5: Adaptive Orchestration the interval loop itself routes work per contact, and the method here is what connects that loop to the planning loops above it.

See Also

References

  1. ↑ Blackburn, J. D., Kropp, D. H., Millen, R. A., "A comparison of strategies to dampen nervousness in MRP systems", Management Science 32(4), 1986, pp. 413–429.
  2. ↑ Sterman, J. D., Business Dynamics: Systems Thinking and Modeling for a Complex World, McGraw-Hill, 2000, chapter 17.
Retrieved from "https://wiki.wfmlabs.org/w/index.php?title=Variance_Routing_by_Horizon&oldid=5652"