Lease Break Probability Model

A lease break probability model scores how likely a tenant is to exercise a contractual break option before the notice deadline. For each lease with a break clause, it returns a probability (and usually a risk band) so asset managers can prioritize outreach, retention offers, and backfill planning while they still control the calendar.

The score is not a renewal decision. It is an early warning input. The asset manager still decides whether to engage, what commercial terms to float, and whether to prepare reletting. The model’s job is to rank where break risk is concentrated and to surface which signals are driving that rank.

The model combines operational, financial, market, and relationship evidence. Typical feature families include:

• Space utilization. Badge, sensor, booking, or occupancy signals that show how fully the demised premises are used relative to lease capacity and to the tenant’s own baseline. Sustained underuse often precedes downsizing or exit at a break date.
• Financial health. Credit ratings, payment performance, cash-flow or covenant indicators, and other credit data that speak to the tenant’s ability and incentive to keep the space.
• Market rent differential. The gap between passing rent and current market rent for comparable space. Tenants paying above market have a clearer economic case to break and relocate or renegotiate; tenants below market may still break for operational reasons, but the rent gap remains a core feature.
• Communication sentiment. Tone and content from emails, meeting notes, broker correspondence, and formal notices when those channels are available. Mentions of hybrid strategy, footprint reduction, relocation, or dissatisfaction add context that utilization and rent alone miss.

Feature engineering should respect lease mechanics: break date, notice period, remaining term, and any conditions attached to the break. A high utilization score six months before a break with a short notice window means something different from the same score three years out.

Training labels usually come from historical outcomes: exercised breaks, non-exercised breaks that ran to expiry or renewal, and early terminations where the break was the contractual path. Where history is thin, start with transparent scorecards and graduate to supervised models only when labeled events are enough to validate calibration.

The model must not invent a probability from partial evidence. If space utilization, financial health, or market rent inputs are missing for a lease, the output for that lease is empty (no score, no band, no ranked position). Sentiment alone is never enough to publish a break probability.

Empty output is deliberate. A number built on two of three pillars looks precise and is easy to misuse in boards and lender packs. Asset managers need a clear signal that the record is incomplete so they can fix data coverage rather than act on a hollow score.

Operational rules that keep this honest:

1. Treat each of the three required families as a gate, not a nice-to-have weight.
2. Log which gate failed so data owners can remediate utilization feeds, credit refresh, or valuation/comps.
3. Keep the lease visible in the renewal pipeline with a “score unavailable” status so it is not silently dropped from human review.
4. Re-score only after the missing family is present and timestamped; do not backfill with proxies that pretend to be the missing input.

Sentiment may still be stored and shown as qualitative context when a score is blocked, but it must not substitute for a probability.

Once a score exists, the asset manager owns the plan. Typical workflow:

1. Sort the break calendar. High-probability leases with approaching notice deadlines move to the front of the week’s worklist.
2. Inspect drivers. Review which features lifted the score: empty floors, deteriorating credit, above-market passing rent, or negative correspondence. Drivers shape the conversation more than the percentage itself.
3. Choose a retention or exit path. Options include early renewal talks, rent or incentive packages, rightsizing within the portfolio, or orderly reletting. The model does not pick the path.
4. Align with commercial drafting. When retention is the path, hand risk context into heads-of-terms and rent strategy workstreams so offers match why the tenant might leave.
5. Track outcomes. Record whether the break was exercised, waived, or renegotiated, and feed that back for calibration reviews.

Human-in-the-loop remains non-negotiable at notice and offer stages. Legal notice, landlord consent conditions, and relationship history are outside the model. The score ranks urgency; people still plan and approve.

Break clauses are sparse events. A portfolio may see only a handful of exercises per year, so probability estimates drift if left unchecked. Recalibrate on a fixed cadence (for example after each major valuation cycle or annually) using holdout leases and recent outcomes. Watch for:

• Overconfidence near 0 or 1. Extreme scores with thin local history deserve wider uncertainty bands or manual override flags.
• Market regime shifts. Rapid rent moves, hybrid-work policy changes, or sector stress can invalidate features trained on quieter periods.
• Leakage from future knowledge. Do not train on post-notice correspondence that would not have been available at the decision time you are simulating.
• Portfolio mix changes. A model tuned on office stock will mis-rank industrial or retail breaks if applied blindly.

Document assumptions about notice windows, cure rights, and multi-break schedules so analysts interpret scores against the same lease rules. When two leases share a tenant, avoid double-counting group-level credit without a clear aggregation policy.

• Automated Renewal Heads of Terms
• Renewal Pipeline Workflow Agent
• Renewal Rent Optimization Model