Red Sea Shipping Disruptions 2024: Impact on Lead Time AI Models and Safety Stock Assumptions

Event Summary

Starting in November 2023, Houthi forces in Yemen began targeting commercial vessels transiting the Red Sea and Bab-el-Mandeb strait. By January 2024, the majority of major container carriers — including Maersk, MSC, CMA CGM, Hapag-Lloyd, and Evergreen — had suspended Red Sea transit and rerouted vessels around the Cape of Good Hope. The rerouting added roughly 3,500–4,000 nautical miles per voyage.

The operational consequence was immediate and measurable. Asia-Europe transit times that had averaged 25–28 days via Suez extended to 35–40 days via the Cape. For Asia-US East Coast lanes that used the Suez route, the extension was 7–10 additional days. Spot freight rates on affected lanes spiked sharply in January and February 2024 before partially stabilizing, though they remained well above pre-disruption levels through most of the year.

Planning Variables Affected

How AI Lead Time Models Failed

Most AI-driven inventory optimization tools calculate safety stock using some variant of the standard formula that multiplies a service-level z-score by the standard deviation of demand and lead time. The lead time component is typically estimated from historical purchase order receipt data — either as a fixed average or as a distribution fitted to past observations.

The Red Sea disruption exposed three specific failure modes in how these models handle structural regime changes:

Failure Mode 1: Lag in Lead Time Distribution Update

Most tools update their lead time estimates based on rolling windows of actual receipt data — commonly 13, 26, or 52 weeks. When rerouting began in December 2023–January 2024, the new 38-day transit times did not appear in receipt data until 5–6 weeks later. Models continued recommending safety stock levels calibrated to 26-day lead times for the entire period between the rerouting decision and the first Cape-routed receipts clearing port.

For a product with 4 weeks of average demand, this lag translated to a safety stock shortfall equivalent to roughly 1.5–2 weeks of cover. The exact shortfall depends on the model's specific safety stock formula and the service level target, but the directional error was consistent across tools.

Failure Mode 2: Mean Reversion Pressure in ML Models

Several machine learning-based lead time forecasting models — particularly those using gradient boosting or LSTM architectures trained on multi-year data — exhibited mean reversion behavior. When Cape-routed receipts began arriving at 38–40 days, the models partially discounted these observations as outliers rather than treating them as the new regime. This is a known behavior in models optimized for accuracy on historical data: they resist large structural shifts because such shifts were rare in training data.

The practical result was that models converged toward a blended lead time estimate — something like 30–32 days — rather than fully adopting the 38-day Cape baseline. Safety stock recommendations remained below what the actual supply situation required.

Failure Mode 3: Variance Underestimation

Even models that updated their mean lead time estimate correctly often underestimated the increase in lead time variance. Cape of Good Hope routing introduced new variability sources: weather delays in the South Atlantic, port congestion at intermediate stops (Port Said, Colombo, Tanjung Pelepas), and carrier schedule instability during the rerouting transition. The standard deviation of transit time roughly doubled during Q1 2024 compared to the 2022–2023 baseline.

Since safety stock requirements scale with lead time variance — not just mean lead time — this underestimation compounded the shortfall from the mean shift. A model that correctly updated its mean to 38 days but kept its variance estimate at the historical 2.5-day standard deviation would still understate safety stock by a meaningful margin compared to a model using a 5–7 day standard deviation.

Corrective Actions Taken by Planning Teams

Planning teams that caught the model lag early — typically those with direct carrier communication or freight forwarder alerts — took several manual override approaches:

• Manual lead time override: Setting a fixed lead time parameter of 38–42 days in the planning system, bypassing the model's data-driven estimate, effective from January 2024.
• Safety stock floor: Establishing a minimum safety stock quantity based on the new transit time, regardless of model output, for all SKUs sourced from Asia on affected lanes.
• Variance buffer add-on: Adding a fixed days-of-supply buffer (typically 5–7 days) on top of the model's recommended safety stock to account for schedule unreliability.
• Carrier-specific lead time tables: Splitting lead time assumptions by carrier where some carriers maintained Red Sea transit under naval escort (e.g., some US-flagged vessels) versus full Cape rerouting.
• Expedite threshold adjustment: Lowering the reorder point trigger to initiate expedite reviews earlier, compensating for the longer replenishment cycle.

Safety Stock Recalibration: What the Math Requires

The standard safety stock formula under demand and lead time uncertainty is:

SS = Z × √(LT × σ_D² + D̄² × σ_LT²)

Where:
  Z      = service level z-score (e.g., 1.65 for 95%)
  LT     = mean lead time (days)
  σ_D    = standard deviation of daily demand
  D̄      = mean daily demand
  σ_LT   = standard deviation of lead time (days)

Plugging in representative numbers illustrates the magnitude of the recalibration needed. For a product with mean daily demand of 100 units, demand std dev of 20 units, and a 95% service level target:

The variance effect is not small. A model that correctly updated its mean lead time to 38 days but kept the pre-disruption variance estimate would still carry roughly 40% less safety stock than the correctly calibrated position. This is why the third failure mode — variance underestimation — is operationally as significant as the mean shift.

Affected Supply Chain Functions

Model Governance Implications

The Red Sea disruption is a clean case study in why AI planning models require human-in-the-loop governance structures that go beyond standard model performance monitoring. The models were not malfunctioning by their own metrics — MAPE and bias statistics may have looked acceptable for weeks after the rerouting began, because the error was not yet visible in receipt data.

The failure was a structural regime shift that the models had no mechanism to detect from internal data alone. Detection required external signal integration — specifically, carrier announcements and freight market data — and a governance process that could translate those signals into planning parameter overrides quickly.

Organizations with formal disruption monitoring processes — where a named owner reviews carrier and port news and has authority to override model parameters — responded faster than those relying on the model to self-correct. The median response lag in the latter group was 6–10 weeks from disruption onset to corrected safety stock levels, based on observations from freight and planning consultancies active in this period.

Current Status (as of Q2 2026)

As of Q2 2026, the security situation in the Bab-el-Mandeb strait has not returned to pre-2024 conditions. A significant portion of Asia-Europe container volume continues to route via the Cape of Good Hope. Planning teams that recalibrated their lead time and safety stock assumptions in 2024 should treat the extended lead time as the current baseline — not as a temporary deviation to be reverted.

Models that were manually overridden in early 2024 and then allowed to revert to data-driven estimates may now be correctly calibrated to Cape-route lead times, depending on how much Cape-route receipt history has accumulated. However, teams should verify that variance estimates have also updated — not just mean lead times — since variance normalization takes longer in rolling-window models.

• Verify that your planning system's lead time estimate for Asia-Europe SKUs reflects 35–40 days, not the pre-2024 Suez baseline of 25–28 days.
• Check lead time standard deviation estimates, not just means. If your model shows σ_LT below 4 days for Cape-routed lanes, the variance estimate is likely still anchored to pre-disruption data.
• Do not assume a return to Suez routing in planning parameters without a confirmed, sustained change in the security situation and carrier announcements.
• Review whether your model governance process now includes an external event trigger for lead time parameter review — if not, the next disruption will produce the same 6–10 week correction lag.