Planning Nervousness and Frozen Zones

From WFM Labs

Planning nervousness and frozen zones describes the cost that re-planning on a rolling horizon has of its own, and the two remedies the planning literature found for it: freezing the near horizon, and charging for changes to it. The page matters to anyone proposing to replace an annual workforce plan with a continuous one, because the obvious objection, that a plan which moves every week is no plan at all, is correct unless the design answers it. The answer is not to re-plan less often. It is to commit only the first step of each plan, hold a frozen zone shorter than the delay it protects, and pay a stated cost to break it.

The page belongs to the series that begins at Workforce Planning as a Control System.

Nervousness, defined

A rolling-horizon plan is recomputed each period on new information. Each recomputation can change decisions that were already in motion: a requisition opened on last month's plan is cancelled on this month's, a class start is moved, a partner's band is renegotiated. The accumulated reversal of decisions is nervousness, or schedule instability. It is not forecast error; a plan can be nervous on a perfect forecast if small changes in the inputs move the optimum across a threshold.

The material requirements planning literature studied it directly. In a comparison of strategies to dampen nervousness, freezing part of the planning horizon and charging a cost for changes dominated the alternatives under most conditions; safety stock and lot-for-lot ordering, the remedies one might expect, were not cost-effective.[1] Freezing the near horizon was found to cost little: freezing up to half the planning horizon had a marginal effect on production and inventory cost, and freezing by order count outperformed freezing by period.[2] The conclusions held under demand uncertainty.[3]

Control engineering reached the same design from the other direction. Model predictive control solves an optimization over a finite horizon at every step, applies only the first decision, observes the result, and re-solves.[4] The far horizon is provisional by design; only the committed step is firm. Continuous planning in this sense commits the near step and treats the rest as provisional, which is the opposite of updating everything all the time.

The design

Three elements, applied at every horizon of the control-system reading.

  • Commit only the first step. The capacity loop's decision this month is this month's hiring release, not next quarter's. Next quarter's number exists in the plan and is expected to change.
  • A frozen zone shorter than the delay it protects. Decisions inside it are not reversed. For the capacity loop that is a fraction of the requisition-to-seat time (a third of the pipeline in the simulation); for the short-term loop, the published week; for the structural loop, the contract notice period. In the companion simulation the closed-loop policy carries a four-week frozen zone on a twelve-week pipeline, a third of the delay, implemented as a hard delay between decision and release with no cancellation inside it; decisions inside the zone are counted as part of the supply line, since they are as committed as an open requisition [estimated].
  • A change cost. Every change to the plan of record is charged, in the simulation at a flat rate per full-time equivalent moved, and the same charge applies to the conventional cascade. The charge is what makes the plan's own stability a term in the objective rather than a complaint about it.

Two further rules follow from the frozen zone. The planning horizon must be extended by the length of the zone, because a decision released four weeks later than it was taken needs to have looked four weeks further ahead. And the persistence test that routes variance to hiring, see Variance Routing by Horizon, adds its own latency, so the horizon extends by that too.

What it costs

A frozen zone makes the plan late by its own length when the world moves. In the simulation that shows as a small cost under a large upward step, where the closed-loop policy waits for the gap to persist and then waits again for the zone to pass, while a policy that hires the whole gap at once is earlier and pays for it in overshoot later. Across a seasonal year with a trend and disruptions the closed-loop policy makes about a quarter fewer hiring moves than the cascade (719 against 943), with about a third of the cancelled requisitions, at roughly three percent lower cost per contact and statistically indistinguishable service [estimated]. These are results for the combined policy; isolated, the persistence and frozen-zone switches cost more than they save under fast upward shocks and with long pipelines, where waiting for a gap to persist costs more than the reversals it prevents. The MRP studies found that freezing up to half the planning horizon had only a marginal effect on cost;[2] how far that transfers to hiring delays is untested.

A plan can also be nervous in a way no frozen zone fixes: when its inputs are nervous. A forecast that swings with every week's actual produces a plan that swings with it, zone or no zone. That is the job of the smoothing in the forecast and of the persistence test, not of the freeze.

Measuring it

Two measures separate nervousness from legitimate movement. The first is the plan-of-record reversal: sample the plan at the decision cadence, and wherever two consecutive moves change direction, count the smaller of the two. Seasonal turns register once; oscillation registers on every half-cycle. The second is the tracking error, the mean gap between the agents on the floor and the plan of record, which measures whether the plan is being followed at all. A plan with low reversals and high tracking error is stable and ignored; one with high reversals and low tracking error is followed and nervous. A cascade anchored to the annual plan that ignores its supply line shows the second pattern once it rings; a closed loop that follows every week's forecast without a frozen zone shows it too, for a different reason.

Limitations

The MRP results were derived for production schedules with lot sizes and setups, not for hiring; the transfer is by structure, and the numbers do not carry. The frozen-zone length is a design choice; the simulation uses a third of the pipeline and the literature supports freezing in general rather than any particular length. The change cost is a modelling device standing in for the real costs of reversal, which are mostly organisational and rarely booked.

Maturity Model Position

Frozen zones exist at every level of the WFM Labs Maturity Model™ in the form of the published schedule and the locked forecast (see Forecast Lock Process); what changes with maturity is whether the zone is designed to the delay it protects or inherited from the calendar. Setting the zone per horizon and charging for changes belongs with Level 4: Planning in Distributions; the plan of record in Long-Term Planning Agents and the Plan of Record is the capacity loop's frozen zone made explicit.

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. ↑ 2.0 2.1 Sridharan, V., Berry, W. L., Udayabhanu, V., "Freezing the master production schedule under rolling planning horizons", Management Science 33(9), 1987, pp. 1137–1149.
  3. ↑ Sridharan, V., Berry, W. L., "Freezing the master production schedule under demand uncertainty", Decision Sciences 21(1), 1990, pp. 97–120.
  4. ↑ Rawlings, J. B., Mayne, D. Q., Diehl, M., Model Predictive Control: Theory, Computation, and Design, 2nd ed., Nob Hill Publishing, 2017.
Retrieved from "https://wiki.wfmlabs.org/w/index.php?title=Planning_Nervousness_and_Frozen_Zones&oldid=5654"