Microwave Vs Fiber Network Links For Cross Market Latency

When to Use

Use this skill when deciding whether a cross-market strategy should ride a licensed microwave path or a fiber path — the Chicago CME Aurora to New Jersey (Secaucus NY4 / Carteret / Mahwah) corridor being the canonical case — and when arbitrating between them in real time as weather degrades the radio link.

The headline is real: radio travels the corridor at essentially $c$, fiber at roughly $2c/3$ over a longer path, and best-in-class published figures are ~8.2 ms round trip on microwave against ~12.98 ms on the fastest purpose-built fiber. But two things routinely turn that headline into a wrong decision:

1. A propagation-only model is a floor, not an estimate. A radio corridor is a chain of repeater towers — 22 on the shortest reconstructed CME–NY4 path — and the last hop into each data centre is fiber, not air. Bhattacherjee et al. rank five competing Chicago–NJ networks within 0.4–8.1 µs of each other and note that if per-tower added latency exceeded 1.4 µs, the ranking between two of them inverts. The equipment term this skill's previous version dropped is larger than the margin the whole exercise is trying to resolve.
2. The trade-off is availability, not speed. McKay Brothers' co-founder said their microwave network "was down 1 per cent of the time during trading hours in December and January," against Spread Networks' claimed 99.999% fiber availability. Two nines against five. A strategy that only clears its costs at 8 ms is adversely selected during precisely the ~1% of trading time when everyone fast is on fiber.

The engine decomposes each link into propagation, repeater and serialization terms, labels a figure that omits the equipment terms as a lower bound rather than presenting it as the link's latency, and returns a routing verdict that fails closed — unknown weather states, stale telemetry and a simultaneously-degraded backup each produce an explicit status instead of defaulting onto the fragile link.

When NOT to Use

• As a latency measurement tool. This module reads no clock and probes nothing. It is arithmetic over a link configuration you supplied. If you want the corridor's actual latency, measure it — see tick-to-trade-latency-measurement and latency-monitoring-percentile-based-slas.
• For the in-host or facility latency budget. Rack-to-rack and NIC-to-NIC budgets belong to colocation-latency-budget-accounting; multi-facility siting and TCO ranking to co-location-provider-selection-and-network-topology. This skill owns only the inter-market link choice.
• On an asymmetric path. *_rtt_ms is 2 × one_way and is therefore valid only when both directions use the same medium. The common cross-market topology is microwave out and fiber back; budget each leg separately.
• With the shipped equipment terms left at zero. repeater_count, per_repeater_latency_us and fiber_tail_km default to zero and this skill supplies no default per-tower latency — that figure comes from your radio vendor's datasheet, and inventing one would be worse than omitting it. Left at zero, the report is a floor and says so.
• As a spectrum, licensing or link-budget design tool. Fixed point-to-point microwave in the US is licensed under 47 CFR Part 101 and requires frequency coordination. Rain attenuation is predicted with ITU-R P.838 (specific attenuation), P.837 (rain rate statistics) and P.530 (link design). This module implements none of them: it consumes a fade/SNR threshold you derived from them.
• As a real-time circuit breaker. This is a stateless evaluator, not a control loop. Halting trading on a link outage belongs in kill-switch-and-drawdown-circuit-breakers.

Prerequisites

• Route distances, not great-circle distances. For CME Aurora to Nasdaq Carteret the geodesic is ~1,176 km; Spread Networks' fiber is 1,328 km of glass, whose ground path Bozkurt et al. measured at 1,253 km with the remaining ~6% attributed to slack coils left for future repairs. Feeding a geodesic into a fiber budget understates it by roughly that whole 12%.
• The fiber type actually lit on the route. propagation_speed_km_s defaults to 204,190 km/s (Corning SMF-28 group index 1.4682 at 1550 nm, ITU-T G.652). G.655 NZ-DSF runs at 203,940 km/s and ultra-low-latency fiber at 205,056 km/s. Over 1,328 km that choice alone is ~8 µs one-way — larger than the margin separating competing radio networks.
• The equipment terms: repeater_count (towers, or in-line amplifiers on fiber), per_repeater_latency_us from the radio/amplifier datasheet, and fiber_tail_km for the data-centre-to-tower fiber at both ends.
• Link telemetry per link: current_weather from a vocabulary the engine knows, packet_loss_pct, signal_to_noise_ratio_db, and ideally telemetry_age_s.
• Thresholds you calibrated: min_snr_db is your radio's modulation-specific demodulation floor plus its fade margin; max_telemetry_age_s is how stale an observation may be before it stops being evidence. Both default to None, which disables the check — set them.
• Fiber telemetry. Optional, and the single most valuable optional input: without it, failover into a dead backup is indistinguishable from a successful one.

Workflow

1. Decompose the latency, and refuse to launder a floor as an estimate.
   • One-way propagation is $\tau = \frac{d_{\text{km}}}{c/n} \times 1000$ ms, with $n = 1.000315$ for radio (ITU-R P.453 average sea-level refractivity $N_0 = 315$) and $n = 1.4682$ for fiber. Round trip is $2\tau$ only on a symmetric path.
   • Add the equipment terms: repeater_count × per_repeater_latency_us, the fiber tail at fiber speed, and serialization if a payload size is given.
   • If no equipment term was supplied, is_microwave_lower_bound_only is True and the audit note says LOWER BOUND. Treat that report as an upper bound on the advantage, never as the link's latency.
2. Compute the advantage on unrounded values. $\Delta = \text{RTT}{\text{fiber}} - \text{RTT}{\text{mw}}$ and $\text{Adv}% = \Delta / \text{RTT}{\text{fiber}} \times 100$. Rounding is applied to the report fields only. Defaulting fiber to the fastest common type (G.652) is deliberate: it understates the microwave advantage rather than flattering it.
3. Classify the telemetry before believing it. A weather string outside known_weather_states, or an observation older than max_telemetry_age_s, is unusable — not fine. Both fail closed to fiber. This is the difference between a vendor renaming an enum and a silent month of routing over an unassessed link.
4. Test the degradation signals, in the order they lead. SNR collapses first, the radio drops to a more robust modulation, and only then do packets disappear. Weather state and packet loss are lagging confirmations; min_snr_db is what catches a fade before it costs fills.
5. Resolve the status by precedence, and treat a lost backup as its own outcome: NO_HEALTHY_LINK_ESCALATE > FAILOVER_TO_FIBER_RAIN_FADE > FAILOVER_TO_FIBER_TELEMETRY_UNUSABLE > HOLD_FIBER_RECOVERY_HYSTERESIS > ROUTE_MICROWAVE_PRIMARY. When both links are degraded the engine nominates no route and returns selected_link_type = "NONE". Stop the cross-market strategy; do not let it pick the least-bad degraded path on its own.
6. Return to radio on hysteresis, never on a single clean reading. Failover trips above max_packet_loss_threshold_pct (default 1.0%); recovery requires loss at or below recovery_packet_loss_pct (default 0.1%) for recovery_dwell_evaluations consecutive evaluations (default 3). A single shared threshold flaps: loss oscillating either side of 1.0% swaps the route on every evaluation, and each swap reorders packets in flight on a path whose entire competitive margin is single-digit microseconds. The engine is stateless — carry previous_status and consecutive_clean_evaluations forward yourself.
7. Read the frequency note as a design signal. ITU-R P.530: rain attenuation "can be ignored at frequencies below about 5 GHz, but must be included in design calculations at higher frequencies, where its importance increases rapidly." A 6 GHz network and an 11 GHz network on the same corridor are not the same product.

Full procedure: see references/workflows.md. Standards reference: see references/standards.md. Printable pre-flight checklist: see assets/checklist.md.

Common Pitfalls

• Quoting a propagation-only figure as the link's latency. It is a floor no physical link achieves. The published measured figure for Aurora→Carteret was 3.982 ms one-way in 2016; propagation over the 1,176 km geodesic is 3.924 ms. The ~58 µs gap is towers, radios and fiber tails — and it is roughly ten times the 0.4–8.1 µs that separates competing networks on that corridor. A model that drops it cannot rank two carriers.
• Assuming an unrecognised weather string means good weather. A vendor enum change, a typo, or an empty field previously fell through to ROUTE_MICROWAVE_PRIMARY. Nothing errors, nothing alerts, and live orders route over a link whose condition was never assessed. Unknown state must fail closed.
• Failing over to a backup you never checked. If the fiber path is also degraded, a "successful" failover is indistinguishable from a working one in every log line it produces. Supply fiber_telemetry and handle NO_HEALTHY_LINK_ESCALATE.
• Using one packet-loss threshold for both directions. Loss hovering around 1.0% flaps the route on every evaluation. Mid-flight route changes reorder packets across paths 5 ms apart in latency — worse for a sequenced feed than staying on the slow path.
• Sizing the strategy on microwave latency alone. At ~99% trading-hour availability, the fiber path carries you for on the order of 1% of the session — and it is not a random 1%, it is the volatile weather-driven 1%. If the edge does not survive at fiber latency, it does not survive.
• Assuming the "microwave" path is all air. Data centres are not on towers. The corridor is hybrid, with fiber tails at both ends that run at $2c/3$; leaving fiber_tail_km at zero understates the radio path by tens of microseconds.
• Ignoring which fiber is lit. G.652 versus G.655 NZ-DSF is ~8 µs one-way over this corridor. Worse, Bozkurt et al. record that "in some instances, fiber providers deliberately increase the amount of fiber in a link (through the addition of fiber spools) for the purpose of service differentiation" — your circuit may be slowed on purpose relative to the premium tier.
• Feeding a great-circle distance into a fiber budget. Spread Networks' corridor is 1,328 km of glass against an 1,176 km geodesic: +12%, and that is the shortest route ever built for the pair.
• Forgetting that bandwidth is a latency term. FCC Part 101 coordinates up to 60 MHz at 6 GHz and 80 MHz at 11 GHz — narrow channels. A single 1,500-byte frame costs 120 µs of serialization on a 100 Mbps radio link against 1.2 µs on 10 Gbps fiber. This is why radio corridors carry a curated subset and the bulk feed stays on fiber.
• Letting a NaN into a distance or a loss figure. Every comparison against NaN is False, so a NaN distance yields a NaN RTT that never breaches anything and renders as a clean primary route. The engine rejects it instead.

Verification

• Constants derived, not rounded: $n_{\text{air}} = 1.000315 \implies 299{,}698.05$ km/s (3.336692 µs/km); $n_{\text{group}} = 1.4682 \implies 204{,}190.48$ km/s (4.897388 µs/km). No medium may be configured faster than $c$ — propagation_speed_km_s = 300000 is rejected as superluminal.
• Corridor arithmetic, hand-derived: 1,176 km of air $\implies$ 3.9239494 ms one-way, 7.8478988 ms RTT; 1,328 km of G.652 fiber $\implies$ 6.5037313 ms one-way, 13.0074626 ms RTT $\implies$ saving 5.159564 ms, advantage 39.67%.
• Lower-bound discipline: propagation-only microwave (3.9239 ms) lands below the published measured 3.982 ms one-way; adding 22 towers at 2.5 µs and a 20 km fiber tail moves the round-trip saving to 4.853668 ms (37.31%) — the equipment term costs 306 µs of the headline advantage, and clears the LOWER BOUND warning from the audit note.
• Fiber type: G.652 versus G.655 NZ-DSF over 1,328 km $\implies$ 7.97 µs one-way.
• Serialization: 1,500 bytes $\implies$ 120.0 µs at 100 Mbps, 12.0 µs at 1 Gbps.
• Fail-closed: 'HEAVYRAIN', '', and telemetry aged 600 s against a 30 s bound each $\implies$ FAILOVER_TO_FIBER_TELEMETRY_UNUSABLE, never ROUTE_MICROWAVE_PRIMARY.
• Boundary: exactly 1.0% loss $\implies$ ROUTE_MICROWAVE_PRIMARY; 1.0000001% $\implies$ FAILOVER_TO_FIBER_RAIN_FADE, with the reason string rendered in significant digits so a marginal breach does not read as 1.000% > 1.000%.
• Hysteresis: from a failover state, four consecutive clean evaluations at 0.02% loss $\implies$ HOLD, HOLD, PRIMARY, PRIMARY; loss of 0.5% (inside the band) $\implies$ HOLD_FIBER_RECOVERY_HYSTERESIS with the dwell counter reset to 0.
• Both links degraded $\implies$ NO_HEALTHY_LINK_ESCALATE with selected_link_type = "NONE" and an empty link id.
• Argument safety: mismatched telemetry.link_id, a fiber config passed as the microwave leg, NaN/Inf/negative/out-of-range inputs, and a recovery threshold not strictly below the failover threshold each raise NetworkLinkError (a ValueError subclass).
• Run python -m unittest discover -s skills/microwave-vs-fiber-network-links-for-cross-market-latency/scripts.

Related Skills

• co-location-provider-selection-and-network-topology
• colocation-latency-budget-accounting
• tick-to-trade-latency-measurement
• latency-monitoring-percentile-based-slas
• network-jitter-impact-on-strategy-performance
• cross-venue-latency-arbitrage-defensive-design
• multi-region-failover-for-broker-connectivity
• kill-switch-and-drawdown-circuit-breakers