Methodology

Version 1.2 · Odoma Tracker · Tallinn, Estonia · odoma.ee

1. Problem statement

Road transport accounts for roughly one quarter of EU greenhouse-gas emissions, with passenger cars responsible for around 61% of road-transport CO₂ (EEA — Greenhouse gas emissions from transport; European Parliament — CO₂ emissions from cars). Technical improvements — electrification, EU 2035 fleet targets (European Commission — Cars and vans) — address the hardware side. Driving behaviour, however, accounts for an estimated 10–30 % variance in fuel consumption for the same car on the same route, depending on acceleration, braking, speed consistency, and idle time.

This behavioural component is largely unaddressed by existing solutions. OBD-based telematics need hardware investment and are rarely adopted voluntarily. Manufacturer in-car systems reach only new-vehicle owners. Navigation apps optimise for time, not for ecological efficiency. Insurance UBI programmes (see §7) cover roughly 13.8 million policies in Europe at the end of 2024 (Berg Insight 2025) — a small fraction of the fleet, concentrated in Italy and the UK.

Odoma targets the gap with a privacy-first, smartphone-native approach.

2. Our approach

2.1 GPS-only architecture

Odoma computes scores using GPS and inertial sensors available natively on any modern smartphone — iPhone or Android — no OBD dongle, no proprietary hardware, no manufacturer dependency.

This is a deliberate trade-off. OBD-connected systems achieve higher per-event accuracy (up to 6 % fuel savings in the EU ecoDriver controlled study) but require per-vehicle hardware installation, capping adoption at a few percent of the fleet. GPS-only systems show lower per-driver precision but enable population-scale deployment across the entire fleet, including older vehicles.

The EU ecoDriver programme (FP7, €14.6 M, 170 drivers across 7 countries) established 2.5 % fuel savings as the benchmark for GPS-only smartphone applications (CORDIS project 288611). Odoma uses this as its primary reference point.

2.2 Real-time + post-trip feedback

Research consistently shows that the combination of real-time feedback and post-trip summary produces larger and more durable behavioural effects than either alone (ecoDriver, 2016; meta-analysis of 17 studies, Sanguinetti et al., 2020).

Odoma provides:

• Real-time eco meter during each trip (speed smoothness, acceleration events).
• Post-trip score (0–100) with component breakdown.
• Historical trends and streaks to reinforce habit formation.

2.3 Gamification for long-term retention

The primary challenge in eco-driving interventions is behavioural fade: improvements typically decay within weeks to months after initial adoption (ecoDriver longitudinal data: 4.6 % → 2.5 % over 10 months on urban roads).

Gamification addresses fade through intrinsic motivation. The GamECAR project (H2020, €1 M) demonstrated 0.59 L/100 km average savings (p = 0.004, Cohen's d = 1.026 — large effect) using eco-score, achievements, leaderboard, and challenge mechanics (CORDIS project 732068).

A 22-month German study (n = 108, BEV drivers) showed gamification builds transferable skill that partially persists after app removal, while financial incentives do not produce the same transfer (Baumgartner et al., 2019).

Odoma implements: eco score per trip, cumulative streak tracking, global leaderboard with vehicle-category normalisation, and fuel/CO₂ loss transparency.

3. Eco score methodology

Revised 22.09.2026 — this section describes the formula that actually runs in production (eco algoVersion 7.1.1, July–September 2026), not the original v1 design. Known weaknesses of the current formula are stated here, not in a footnote.

3.1 Signals

Each trip receives a score from 0 to 100. The score is built from seven behavioural signals derived on the device from GPS speed (1 Hz) and, where available, altitude; no OBD or accelerometer stream enters the score:

How acceleration is measured. Longitudinal acceleration is the difference of consecutive GPS-derived speeds after a 3-point median filter — effectively an average over a ~2-second window, not an accelerometer peak. A threshold of 3.3 m/s² on this signal fires when speed changes by roughly 25 km/h within 2 s; a short, sharp jab that an accelerometer would register as 0.3 g can be smoothed below it. Industry harsh-braking thresholds of 0.20–0.45 g (NJIT — Harsh braking as a crash-risk surrogate, MDPI 2025) are measured on accelerometer peaks and are therefore only nominally comparable with ours.

3.2 Detection pipeline

Detection is deterministic and runs entirely on-device. The trip's GPS samples (typically 1 Hz, 4 m horizontal accuracy) are processed in five passes:

1. Median smoothing. A 3-point sliding median is applied to GPS-derived speed. This removes single-sample spikes from gear changes (manual transmissions) and brief jitter fixes — without altering raw distance.
2. Pass A — bearings. Bearing computed per sample for samples with displacement ≥ 2 m, used downstream for turn detection.
3. Pass B — jam flags. A sample is flagged inJam when ≥ 3 stop-go transitions (< 5 km/h → > 10 km/h) occurred in the last 30 s and current speed is < 15 km/h. Harsh events inside a jam are weighted 0.0 (acceleration) and 0.3 (braking) because stop-and-go is not under voluntary control.
4. Pass C — brakingToStop flags. Forward-looking: does the current deceleration reach a full stop (< 0.3 m/s) within 5 s, and does that stop last ≥ 8 s? If yes, the brake event is weighted 0.5 — it is a stop at a light or a sign, not panic braking.
5. Pass D — accelFromStop flags. Backward-looking mirror of Pass C: was there a near-stop (< 5 km/h, ≥ 5 s) in the last 5 s before this acceleration sample? If yes, the harsh-accel is weighted 0.25 — it is a normal launch from a light, not aggressive driving. (v5.4.1, May 2026: thresholds relaxed from 1 km/h / 8 s to 5 km/h / 5 s after field-testing on Tallinn taxi data, where rolling stops and short light cycles were systematically missed.)

Per-event weights cascade from full-discount to no-discount:

acceleration: jam → 0.0 ; on-ramp → 0.0 ; sharp turn → 0.0 ;
              accelFromStop → 0.25 ; moderate turn → 0.25 ;
              else → 1.0
braking:      jam → 0.3 ; brakingToStop → 0.5 ; else → 1.0

Since September 2026 one manoeuvre counts as one event regardless of the GPS sample rate (the event is registered on entry into the manoeuvre); previously iPhone counted every sample above threshold.

3.3 Grade compensation

When the road grade exceeds ±3 %, thresholds are scaled by 1.4×: uphill multiplies the acceleration threshold (a heavier press is normal climbing the hill); downhill multiplies the braking threshold (engine-braking and gravity require firmer pedal). Grade is derived from successive GPS altitudes, smoothed.

3.4 From signals to score

Each penalty is a share of opportunities (starts, deceleration cycles, manoeuvres), not a count per kilometre, so a short trip is not punished for a single event: one start and one stop per trip are always free. Shares are damped by sample confidence (a share computed from fewer than four starts counts proportionally less). Weights: rough starts 22, rough decelerations 26, smoothness 16, jerk 10, redundant manoeuvres 8, idling 8, speeding 15 (caps). Bonuses are bounded (coasting ≤ 6, cruising ≤ 5, smooth-start series 3).

The raw score (100 − penalties + bonuses) is then blended with a neutral anchor of 85 according to trip confidence (0.45 for trips with fewer than three stop-go events or under 1 km; 0.65 for 3–6 events; 0.85 for seven or more events or 10 km+; lower with poor GPS quality). Because of the blending, a trip on which nothing was detected scores 92, 95 or 98 depending on confidence, and 10 raw penalty points cost 4.5–8.5 score points.

Known limitation (internal audit, 22.09.2026). On the current formula, drivers with 20+ trips cluster between 87 and 95 (standard deviation ≈ 2), while their calculated excess CO₂ from driving style differs by more than 20×. In the synthetic reference corpus, a trip with three full-throttle starts scores the same 95 as a clean one, and three emergency brakings score 84. The scale is compressed at the top: it separates good from very good far better than good from bad. A recalibrated formula (v8) is being prepared with a continuous energy-based term and population-anchored scaling, targeting a median of 75–80; it will run in shadow mode on real drivers before replacing the score, and history will be re-scored so that no cut-over appears as a behaviour change. Until then the public score should be read as a within-driver progress signal, not as a ranking between drivers.

3.5 Driver-loss model

Beyond the summary score, Odoma computes driverLoss — the estimated excess CO₂ emitted because of driving style relative to the vehicle's baseline consumption:

driverLoss (g CO₂) = harshAccelLoss + harshBrakeLoss + idleLoss − coastingBonus

Each component is a calibrated function of the corresponding event count (after context-adjustment) and the vehicle's declared CO₂ norm in g/km. The figure is accumulated across all trips and reported on the public Impact dashboard as totalDriverLossKg — a sum of these calculated estimates, not a direct measurement of emissions.

3.6 Vehicle normalisation

Leaderboard rankings are normalised by vehicle class derived from declared CO₂ norm and fuel type, ensuring a heavy-diesel-SUV driver competes within their class rather than against a BEV driver. Classes: BEV / Eco (≤ 110 g/km) / Compact (≤ 150) / Standard (≤ 185) / Heavy (> 185).

4. Safety scoring

In v1.2 Odoma adds a separate safety score, drawn from the same event stream but with different weights — calibrated to the literature on harsh braking as a crash-risk surrogate (Connected-Vehicle Telematics study, NJIT/MDPI 2025).

Penalties per 100 km of trip distance:

Harsh braking is the dominant predictor of crash risk in the published literature, which is why it carries twice the weight of acceleration in the safety component. Eco scoring weights them roughly equally, because both equally reflect non-economical driving.

The safety score is provisional in v1.2 — it is collected and computed but currently hidden in the UI (as noted in the app's release notes) until calibrated against independent crash data.

5. Behavioural change measurement

5.1 First-week baseline

For each driver, Odoma records the average eco score across their first 7 days of active use as firstWeekAvg. This is an individual behavioural baseline established before any feedback-driven learning has occurred.

5.2 Improvement metric

avgImprovement = mean(last30DaysAvg − firstWeekAvg)

across drivers with ≥ 20 scored trips on the current formula version and tenure over 37 days (so the first week and the last-30-days window never overlap), synced within the last 30 days. Both terms are simple per-trip means computed on the device in the same way; the lifetime average is not used here because it is distance-weighted and would compare unlike quantities. This is a within-subject measure — it compares each driver only to themselves.

The figure is published on /impact as proven only when both hold: n ≥ 30 drivers and the lower bound of the 95% confidence interval is above zero. Until then the page states plainly that we are still measuring, with the current n. (Revised 22.09.2026 after an internal audit: the earlier gate published a +0.1 average whose 95% CI was [−1.5; +1.7].)

5.3 Why this metric is conservative

• Only drivers with ≥ 20 trips are included (eliminates trial users), on one formula version (no mixing of scales across app releases).
• Profiles produced by development or store-screenshot builds are excluded from all statistics (stats_excluded).
• No regression-to-mean correction is applied.
• No control-group comparison — the figure does not claim causality, only within-user change.
• GPS-only accuracy limitations are acknowledged (see §9).

6. Comparison to prior research

Odoma's approach corresponds most closely to the ecoDriver GPS-only condition with the addition of GamECAR-style gamification mechanics. Results reported in the literature for comparable programmes: 2.5 – 6.6 % fuel-efficiency improvement per driver. This is not a forecast or a measurement of Odoma Tracker; our own result is published on /impact.

7. Insurance and usage-based-insurance (UBI) context

Insurance telematics is the largest commercial application of the same underlying signals (acceleration, braking, mileage, time-of-day). Berg Insight reports approximately 13.8 million UBI policies in Europe at end-2024, expected to reach 20.1 million by 2029 (CAGR 7.8 %, Berg Insight 2025). Mordor Intelligence projects the European insurance-telematics market at USD 0.97 B in 2025, growing to USD 2.78 B by 2030 (market report).

Leading European telematics suppliers cited in PTOLEMUS 2025 UBI ranking include IMS, Octo Telematics, The Floow, Cambridge Mobile Telematics, Targa Telematics, Munic, Geotab, Vodafone Automotive, Arity, FairConnect, Dolphin Technologies and Redtail Telematics.

Odoma is not an insurance product. However:

• The eco / safety pipeline collects signals comparable to those used in commercial UBI.
• We publish the methodology in the open so that fleet operators, insurers, and researchers can audit it.
• A separate opt-in, two-tier crowd-sourced dataset (schema documentation) allows aggregate, k-anonymised, GDPR-compliant data sharing with research and pilot partners. Tier A is non-longitudinal and anonymous by default; Tier B requires explicit consent for longitudinal studies. Coordinates and exact timestamps are never uploaded.
• Pilot partnerships with insurers, fleet operators, and public-sector environmental programmes are open — see About → Get involved.

The European Transport Safety Council (ETSC 2019 — Telematics and Road Safety, local mirror) provides the canonical EU-policy reference for how telematics signals translate into measurable road-safety outcomes.

8. Privacy and data architecture

• No GPS coordinates stored on server. All route data remains on device.
• Each user is identified by a randomly generated UUID stored locally (not linked to any personal identifier).
• Server receives only aggregate metrics: avgScore, tripCount, currentStreak, totalKm, wastedCo2Kg, fuelType, country code (ISO 3166-1).
• Leaderboard participation is opt-in. Default is off.
• Country derived from Locale.current.region — not from GPS.
• The crowd-sourced uplink (§7) is separately opt-in and bucketed: speeds, durations, and event rates are reported as discrete bins, never as raw streams. K-anonymity guard auto-promotes small countries (EE / LV / LT / LU / CY / MT) to a higher k-threshold to prevent re-identification.
• Data processing complies with GDPR Article 6(1)(a) (consent), Article 5 (purpose limitation, data minimisation), and Article 25 (privacy by design and by default).
• Full details in our Privacy Policy.

9. Limitations

GPS accuracy. GPS-derived speed and position introduce noise, particularly in urban canyons, tunnels, and at trip start before satellite lock stabilises. Odoma applies median smoothing, jitter floors, and gap-fill, but accuracy is inherently lower than OBD-sourced data. Score comparability across different device hardware is not yet validated.

Sample size. Current cohort is sufficient for directional conclusions but insufficient for statistically robust population-level claims. A target of 500+ drivers across 3+ countries is required for publishable research.

Causal attribution. The improvement metric measures within-user change over time, not a controlled experiment. Confounding factors (seasonal variation, route changes, vehicle changes) are not controlled.

Self-selection bias. Opt-in users are more motivated than the general driving population, likely producing higher improvement rates than a mandatory programme would achieve.

Driver attribution. When the phone-holder is a passenger (taxi, public transport, or driven by a family member), eco / safety scores no longer reflect the holder's behaviour. Odoma classifies non-self trips with three confidence tiers and excludes them from aggregates and from the crowd-sourced uplink — see the passenger-mode design for details.

10. References

1. ecoDriver Consortium (2016). Final Report. European Commission FP7 Project 288611. cordis.europa.eu/project/id/288611
2. GamECAR Consortium (2018). Results Summary. European Commission H2020 Project 732068. cordis.europa.eu/project/id/732068
3. Sanguinetti, A., Queen, E., Yee, C., & Akanesuvan, K. (2020). Average impact and important features of onboard eco-driving feedback: A meta-analysis. Transportation Research Part F: Traffic Psychology and Behaviour, 70, 80–90. doi.org/10.1016/j.trf.2020.02.011 · open-access at escholarship · local mirror
4. Baumgartner, M. et al. (2019). Long-term effects of gamification on eco-driving behaviour. Energy Research & Social Science, 57. doi.org/10.1016/j.erss.2019.101237
5. ETSC (2019). Telematics and Road Safety. European Transport Safety Council. etsc.eu · local mirror
6. Evaluating Harsh Braking Events as a Surrogate Measure of Crash Risk Using Connected-Vehicle Telematics. MDPI Vehicles, 8(3), 68 (2025). mdpi.com
7. European Environment Agency (2024). Greenhouse gas emissions from transport in Europe. eea.europa.eu
8. European Environment Agency (2025). CO₂ performance of new passenger cars. eea.europa.eu
9. European Commission. Cars and vans — climate action. climate.ec.europa.eu
10. Berg Insight (2025). Insurance Telematics in Europe and North America, 9th edition. berginsight.com
11. PTOLEMUS Consulting Group (2025). UBI solution provider ranking. ptolemus.com
12. Mordor Intelligence (2025). Europe Insurance Telematics Market Size. mordorintelligence.com
13. Regulation (EU) 2016/679 — General Data Protection Regulation (GDPR). eur-lex.europa.eu

For partnership enquiries: Contact usCopy Source code, privacy policy, and live impact data: odoma.ee