Methodology — Crisis Analog Engine

How Sirius matches current market patterns against 11 historical crises

Every week, the Sirius model scores the current freight pressure pattern against a library of 11 historical crisis events using cosine similarity across four dimensions. The best-matching analog drives the forward projection scenarios and sets the confidence band on the outlook.

Contents
  1. Why historical analogs
  2. The 11 crises in the library
  3. Four scoring dimensions
  4. How cosine similarity works
  5. Analog confidence scores
  6. How analogs drive forward projections
  7. Why analogs are bounded
  8. Limitations
4-axis pressure pattern match - illustrative example
Fuel Capacity Surcharge FX -- --
Current pattern (W21 - illustrative)
Best analog match (illustrative)
Values and shapes are illustrative. Actual scoring uses normalized pressure readings from the Sirius model.
01

Why historical analogs

Freight markets are not random. They follow patterns that have recurred across crises: a fuel shock builds capacity pressure, capacity pressure drives surcharge increases, surcharge increases tighten the FX impact on cross-currency lanes. The sequence and relative intensity of these four dimensions varies by crisis type, but the structural relationships between them are recognizable.

This means that when a new crisis begins to build in the data, it often resembles a historical precedent - not in its exact magnitude or geographic origin, but in its pressure pattern across the four dimensions. A fuel-led chokepoint disruption will leave a recognizable four-axis signature. A demand-led capacity crunch will leave a different one.

The crisis analog engine exploits this structural similarity. Rather than extrapolating from recent trend alone - which is unstable in the early stages of a crisis - it identifies which historical event most closely resembles the current pattern and uses the trajectory of that event as a calibration anchor for the forward projection.

02

The 11 crises in the library

The analog library contains 11 historical crisis events, each encoded as a time-series of four-dimensional pressure readings normalized to the same base scale. Five events are named in public Sirius materials; the remaining six are referenced as additional calibration events to provide coverage of different crisis types without creating public attribution obligations for events that may be politically sensitive.

2020-2021
COVID-19 port congestion and container shortage
2021
Suez Canal blockage (Ever Given)
2022
Russia-Ukraine war fuel and capacity shock
2024
Red Sea Houthi escalation and route diversion
2026
Strait of Hormuz partial restriction
Additional crises studied
6 further crisis events across fuel shocks, capacity cycles, and chokepoint restrictions

Each crisis in the library is encoded as a weekly time series starting from the first week of measurable pressure elevation through to recovery - defined as all four dimensions returning within 10% of their pre-crisis baseline. This produces a crisis "trajectory" with a shape, a duration, and a resolution pattern for each event.

On crisis duration: The library shows significant variation in crisis trajectory length. The 2021 Suez blockage resolved in days for the physical disruption but took 6-8 weeks for capacity pressure to normalize. The 2020-2021 COVID congestion took over 18 months to fully resolve. This variance in resolution timing is a key input to the analog engine's confidence band calculation.
03

Four scoring dimensions

Each crisis event - both historical and current - is represented as a four-dimensional vector at each week in its trajectory. The four dimensions map directly to the Sirius pressure model layers:

  • Fuel dimension: Normalized fuel pressure reading, capturing the combined effect of bunker, jet fuel, and road diesel market moves relevant to the lane being analyzed.
  • Capacity dimension: Normalized capacity pressure, capturing vessel utilization, container availability, and blank sailing data where available for the crisis period.
  • Surcharge dimension: Normalized surcharge pressure, capturing carrier-declared emergency surcharges, war risk premiums, and congestion charges as a proportion of base rate.
  • FX dimension: Normalized currency pressure, capturing the trade-weighted FX movement against the primary currency pair for the trade lane being scored.

Normalization maps each dimension to a 0-1 scale relative to the observed range across all 11 historical crises. This prevents a crisis with extreme fuel pressure from dominating the similarity score simply because of scale differences.

04

How cosine similarity works

Cosine similarity measures the angle between two vectors in multidimensional space. A score of 1.0 means the two vectors point in exactly the same direction - the same pattern of pressure across all four dimensions. A score of 0 means they are completely orthogonal. A score below 0 means the patterns are moving in opposite directions.

For the crisis analog engine, cosine similarity is calculated between the current week's four-dimensional pressure vector and each week in each historical crisis trajectory. The search is run across the full trajectory of all 11 crises, looking for the point in each crisis where the pattern most closely resembled what is seen today.

The result is a ranked list of (crisis, week-in-crisis) matches with their similarity scores. The top-scoring match becomes the primary analog. The second and third matches are retained as secondary analogs and used to calculate the confidence band: when the top-3 analogs agree closely on forward trajectory, the confidence band is narrow. When they diverge, the band is wider.

Illustrative example: Suppose the current week's pressure vector is Fuel=0.72, Capacity=0.41, Surcharge=0.58, FX=0.29 (illustrative values, not real readings). The analog engine finds that Week 4 of the 2024 Red Sea escalation had a vector of Fuel=0.69, Capacity=0.44, Surcharge=0.61, FX=0.27 - a cosine similarity of 0.98. The engine would then use weeks 5 through 12 of the Red Sea trajectory as the primary analog path for the forward projection.
05

Analog confidence scores

The analog confidence score is a composite measure of how well the current pattern is matched and how much agreement there is among the top-3 analogs about where the pattern is heading. It is published alongside each weekly Sirius reading in the Pro report.

The confidence score has four inputs:

  • Match quality: The cosine similarity of the primary analog. Higher similarity means more confidence that the current pattern resembles a historical precedent.
  • Trajectory agreement: The degree to which the top-3 analogs agree on the direction of the next 4-8 weeks of movement across all four dimensions.
  • Analog diversity: Whether the top-3 analogs are from distinct crisis types or all from the same event. Diverse top-3 results indicate the current situation may be a hybrid not well represented in the library.
  • Lane specificity: Whether the matched analog involved significant pressure on the specific lane being analyzed. A global crisis may match the aggregate pattern but be less informative for individual lanes.

A high-confidence analog reading (above 0.85 on the composite score) is treated as a strong anchor for the forward scenario. A low-confidence reading (below 0.55) flags that the current pattern does not closely resemble any historical precedent - which is itself informative. It suggests either an early-stage crisis that has not yet developed its signature, or a genuinely novel event that the library does not cover.

06

How analogs drive forward projections

When the analog engine finds a high-confidence match, the forward trajectory of that analog becomes the primary input to the Sirius base case scenario for the matched lane. The analog trajectory is scaled to account for the difference in absolute magnitude between the historical event and the current situation, then extended forward from the current matched week in the historical crisis.

The base case scenario represents the analog's most likely trajectory: how long the pressure remained elevated, when the first signs of resolution appeared, and the shape of the decay. The escalation scenario is drawn from the analogs with slower, more extended trajectories in the top-3 set. The de-escalation scenario is drawn from the analogs with faster resolution.

This structure means the scenarios are not arbitrary. They are grounded in observed historical outcomes on the closest precedents available. The uncertainty band reflects genuine historical variance in how similar crisis patterns have resolved, not a uniform percentage band applied mechanically around the base case.

07

Why analogs are bounded and not extrapolated

The analog engine explicitly does not extrapolate beyond crisis resolution. Once the trajectory of the matched analog reaches its resolution point - when all four dimensions return to within 10% of the pre-crisis baseline - the forward projection terminates and hands off to the Sirius baseline model.

This is a deliberate design choice. The further a projection extends beyond the analog's resolution window, the less the historical precedent constrains it. A 16-week forward projection based on a crisis that resolved in 12 weeks is anchored for only the first 12 weeks. After that, any trajectory is equally plausible and the model would be manufacturing false precision.

The Sirius Pro report shows the analog's resolution window explicitly so users can see where the historically-anchored projection ends and where the baseline model takes over. The confidence band widens significantly at and beyond the resolution boundary.

On novel crises: When a new crisis type appears - one with no close historical analog - the engine will correctly report low confidence. This is the appropriate response. It means users should rely more heavily on the current-week reading and apply wider scenario bands than the model would generate for a well-matched historical precedent.
08

Limitations

The analog engine is a useful calibration tool, not a prediction mechanism. Several limitations are important to hold in mind when interpreting analog-anchored projections:

  • The library is finite: 11 crises covering roughly 6 years of modern freight market dynamics. A genuinely novel event - a crisis type not represented in the library - will produce low-confidence matches across all 11 precedents. This is a signal to use wider judgment, not to ignore the reading.
  • Historical context differs: Even a high-similarity pattern match does not guarantee that the future trajectory will follow the historical precedent. Market structure, carrier concentration, regulatory environment, and geopolitical context all change over time. The 2020 container shortage and a hypothetical 2027 container shortage with the same pressure pattern may resolve differently for structural reasons the model does not observe.
  • Four dimensions are not the full picture: The analog engine uses four dimensions. Real crisis dynamics involve many more variables. The engine captures the dominant signals but may miss nuances in demand patterns, political resolution mechanisms, or infrastructure changes that would influence the trajectory.
  • Analog confidence is not probability: A high confidence score means the current pattern closely resembles a historical precedent. It does not mean the outcome will follow that precedent with high probability. Both should be understood as directional guidance, not precise forecasts.

The Sirius index publishes the primary analog name, its similarity score, and the confidence composite for every weekly reading in the Pro report. This transparency allows users to assess the quality of the historical anchor for themselves rather than accepting the projection as a black-box output.

See the current analog match and confidence score

Sirius Pro includes the full analog engine output: primary and secondary matches, similarity scores, and the 8-week forward scenario with confidence band.

Download Sirius Free →