A route-evaluation engine for airline network planning, built entirely on open government data. One config-driven pipeline turns raw BTS, StatCan, BEA, and EIA files into LAUNCH / MONITOR / PASS verdicts for new airline routes, with every demand number flagged observed or modeled and every assumption traceable to a source.
Live dashboard: https://network-lab-tweqoqjxhfrhj7imgmuftl.streamlit.app -
click through the route screen, per-market scenario economics, validation, and
backtest. Everything is reproducible from public sources with make data && make all.
Three studies run through the same engine. The point is not that the model says yes; it is that the model says yes, no, and "watch it" for defensible, traceable reasons, and admits what it cannot see.
1. Porter at Toronto Pearson (YYZ): 5 LAUNCH / 3 MONITOR / 71 PASS. The five LAUNCHes (Washington, Chicago, Boston, Philadelphia, Atlanta) are all large, anchor-backed markets that the engine enters at modest share against real incumbents, then right-sizes down to 3x or 14x weekly. This is the low-cost entry logic of taking a slice of a big market rather than owning a small one. Each is a market Porter does not serve nonstop from YYZ today (verified against Porter's actual T-100 network; the one edge case, Chicago, is documented on the page).
2. WestJet transborder from Calgary (YYC): 0 LAUNCH / 0 MONITOR / 76 PASS. The flagship study recommends nothing, and that is the finding. Every viable unserved YYC transborder market is either already served by WestJet (New York, the California/Nevada/Arizona leisure markets, Houston, the Florida sun routes) or too thin or too thoroughly held by a US carrier's connecting hub (Dallas at 23% share against American). This is consistent with WestJet directing its recent transborder growth to Edmonton and Vancouver, not to more YYC breadth. A model that only ever says yes is a sales tool, not an analysis.
3. Two route post-mortems, same market, two failure modes. Flair and Lynx both flew YYC-LAS into WestJet's fortress hub and both exited. A demand screen ranks the market in the top 17% and would have green-lit both. But Flair sustained the route for two and a half years at a 0.59-0.70 load factor against WestJet's 0.90 and made a route-level economic cut, while Lynx flew one year at 0.53 and died in a corporate shutdown that took every route at once. Same market, one economic verdict and one balance-sheet death. That distinction is the whole argument for why a demand-first screen needs an economics and competition layer, and it is invisible to the demand rank alone.
Every candidate market runs the same decision tree. Each threshold lives in config/assumptions.yaml with a source, a confidence grade, and a sensitivity range.
| Step | Rule | Threshold (assumptions.yaml) |
|---|---|---|
| Candidate filter | metro population floor, distance in range, not already served by the carrier (metro-level, from T-100), real airport, not a satellite of a bigger one | candidate_min_metro_pop 400k, served_market_min_deps_yr 26, candidate_min_airport_seats_yr 250k, satellite_airport_rule 60mi/5x |
| Demand | observed DB1B where it exists; else the market's own 2018 StatCan actual x T-100 corridor growth; else gravity x transfer, capped at the survey's own floor for unanchored markets | transfer_factor (median 0.80), unanchored cap 4,000 x growth |
| Share | QSI-lite vs one-stop competitors reconstructed from T-100 frequencies | qsi_weights, qsi_elapsed_time_exponent |
| Right-size | pick the frequency and gauge that maximize annual contribution at a feasible load factor | min_feasible_load_factor 0.50 (fallback: keep the highest-LF option) |
| Economics | 3 fare x 3 fuel scenario grid, fully-allocated margin and break-even load factor | hurdle_margin_pct 8, fare_scenario_grid_pct, fuel_scenario_grid |
| Verdict | PASS if base margin < 0; LAUNCH if margin >= 8% and the fare-downside stays positive and break-even LF <= 0.85; else MONITOR | hurdle 8%, BELF cap 0.85 |
| Competition guard | a modeled share above 70% with zero nonstop incumbents caps LAUNCH to MONITOR (the one-stop reconstruction probably missed real competition) | > 0.70 share, 0 incumbents |
So Alaska's lone MONITOR (Fayetteville-Springdale-Rogers) is fully traceable: its best-fit margin is +0.2%, positive but below the 8% hurdle, so it is held for watching, not launched or rejected.
Every market carries a 3 fare x 3 fuel contribution-margin grid, so a verdict is read against its own downside, not a single point estimate. The example below is a thin Porter candidate that is negative in all nine cells, a clean PASS the tool does not dress up:
- Reconciliation. Computed T-100 transborder totals match StatCan published figures within 1% nationally and at YYZ/YVR/YYC/YUL (the definitional relationship is derived before comparing; see docs/reconciliation.md).
- Share model against truth. At SEA, where DB1B carrier shares are observed, the QSI-lite share model lands at MAE 7.0 share points across 503 market-carrier rows, reported by market structure in docs/validation.md. The Alaska SEA study exists to run the whole chain where demand, share, and fare are all observable: it resolves to 48 PASS / 1 MONITOR, a negative check that the engine does not invent opportunity at a mature hub.
- Transfer factor. Anchored to the discontinued 2018 StatCan transborder city-pair survey: median 0.80, IQR [0.55, 2.02] across 87 pairs, with the dispersion reported everywhere the factor is used.
- Launch backtest. 48 launched-and-resolved transborder routes (2021-2025) scored by a pre-2022 model: survivors and ceased routes both sit in the top demand decile (0.89 vs 0.86 median percentile), and the gap barely moves when Lynx's shutdown casualties are excluded. Demand rank alone does not predict survival, and the report says so rather than overclaiming a signal that is not there.
The engine implements standard airline-planning methods on open data rather than inventing new ones:
- Gravity demand. A log-linear gravity model, the most widely used method for air passenger demand forecasting, calibrated on 6,002 US city-market pairs (2018-19 vintage; 6,092 in the 2023-24 fit) so its coefficients read as elasticities (gravity models for air passenger volume estimation, Grosche et al.).
- QSI-lite share. A Quality of Service Index share model weighting nonstop over connecting service and discounting frequency, the industry-standard frequency-share approach (Cirium, OAG).
- Spill and economics. Truncated-normal spill to expected boardings and a RASM-vs-CASM contribution build with a break-even load factor, per standard airline route economics (MIT 16.75J airline management).
The hard part is not the methods; it is that there is no public Canadian domestic
O&D, no transborder fare data, and the transborder O&D survey froze in 2018. So
the model is calibrated where truth exists (US markets), transferred to Canada-US
markets with an anchored, dispersion-reported correction factor, and validated
end to end on a US hub. Every demand number carries an observed/modeled flag,
every non-derived value lives in one YAML file with source and sensitivity, and a
proposed nonstop is never scored against an empty market: competitor one-stops
are reconstructed from T-100 first.
ingest/ one client per source -> data/raw/ (gitignored), typed Parquet
warehouse/ DuckDB marts: dim_airport/carrier/metro, fact_segment (T-100),
fact_od_market (DB1B), fact_costs (Form 41), fact_fuel (EIA),
reconcile.py checks computed vs published totals in writing
models/ catchment -> demand -> competition -> QSI-lite share -> spill ->
economics (3x3 grid) -> screen (right-sized LAUNCH/MONITOR/PASS)
backtest/ score real 2021-2025 launches with a pre-2022 model; post-mortems
reports/ generated markdown business cases and ranked tables
app/ Streamlit dashboard reading only the precomputed outputs
No single open-source project combines DB1B, T-100, Form 41, gravity, QSI, and
route economics end to end, which is what this repo is. It is config-driven: a
new study is a YAML file (config/studies/), not a code change. Continuous
integration runs the sanity and consistency tests on every push
(.github/workflows/ci.yml); the consistency tests
fail the build if any reader-facing number or verdict count drifts from the
computed artifacts.
make setup # venv + dependencies (~1-2 min)
make data # download sources, build warehouse, reconcile (~50 min cold, ~1 s cached)
make all # models + backtest + reports (~40 s)
make test # sanity + consistency tests (~4 s)
make app # Streamlit dashboard (reads precomputed outputs)
No API keys required. Raw downloads stay out of git; the hand-compiled reference
files (data/reference/: the launch register and the incumbent-network audit
file) are committed. Deploy steps in docs/deploy.md; demo
storyboard in docs/demo_script.md.
See docs/LIMITATIONS.md: the limitations are stated with the design decision each one forced, including the single-vintage backtest lookahead, the airport-fee proxy, the frequency/gauge right-sizing scope, the unanchored-demand cap, and the fact that the served-market exclusion trusts the T-100 join (cross-checked by the committed incumbent-network file). Planned extensions (per-vintage backtest refit, connectivity view, full Porter business cases, ACS income with a key) are listed there rather than implied to exist.
BTS TranStats (public domain), StatCan (Open Government Licence - Canada; tables 23-10-0253/0256/0249/0255/0257/0259/0312, 17-10-0135), BEA (public domain), EIA (public domain), Census delineation files (public domain), OurAirports (public domain).