What do users complain about in Indonesian e-wallet apps?

Complaint trends in 201,860 Google Play reviews of GoPay, OVO, DANA and ShopeePay, Jul 1 – Oct 3, 2026.

• Model on Hugging Face
• Code on GitHub

Findings

Data: 201,860 Google Play reviews (Jul 1 – Oct 3, 2026). Complaints = PRIMARY (IndoBERT + "star 1–2 OR model"; gold complaint F1 0.922 [0.89, 0.95]). Complete Monday–Sunday weeks Jul 6 – Sep 21 (12 weeks); the Sep 28 week is partial (5.7 days). Dates in UTC (WIB = UTC+7). 95% CIs: Wilson for shares, Newcombe for differences. Numbers: reports/phase7_trends.json.

1. OVO is the most complained-about app by far: over complete weeks, 78.4% [77.3, 79.4] of OVO reviews are complaints, vs 37.9% [37.6, 38.2] for DANA, 24.9% [24.5, 25.3] for GoPay and 12.5% [12.1, 12.9] for ShopeePay (all reviews, duplicates included; phase7_weekly_complaint_share.png).

2. The spikes are single-day events, not lasting shifts: DANA's late-September surge is two separate days, Sep 25 (12,023 reviews, 72% of that week, 83.5% [82.9, 84.2] complaints) and Oct 1 (8,707 reviews, 85.9% [85.1, 86.6]), with normal volume (~790–950 reviews/day) on the other days (phase7_spike_daily.png).

3. DANA's Jul 20 spike was about reward points: "Reward points cannot be redeemed" grew from 1.3% to 14.1% of DANA's complaints that week (+12.8 pp [11.8, 13.8], 724 unique complaint texts), followed by "Cannot pay or transfer" (+10.2 pp [9.2, 11.2]) (phase7_spike_topics.png).

4. DANA's Sep 25 and Oct 1 spikes were about access and payments: "Network error or lag despite good signal" gained the most in the week of Sep 21 (+7.7 pp [7.1, 8.4], 1,005 texts) and "Cannot pay or transfer" in the partial week of Sep 28 (+7.7 pp [7.0, 8.5], 854 texts) (phase7_spike_topics.png).

5. GoPay's Jul 28 spike was an outage day: 4,357 reviews in one day with 81.2% [80.0, 82.4] complaints; the broad outage topic grew most (+28.5 pp [26.8, 30.3]), and among specific topics "Cannot pay or transfer" (+5.2 pp [4.3, 6.2]) (phase7_spike_daily.png, phase7_spike_topics.png).

6. ShopeePay's complaint share doubled without a volume spike for two consecutive weeks (one partial): 23.1% [21.6, 24.7] in the week of Sep 21 (2,703 reviews, normal volume) and 22.1% [20.5, 23.7] in the partial week of Sep 28, vs a median of 11.2% for Jul 6 – Sep 7; two weeks are too few to call it a lasting change (phase7_weekly_complaint_share.png).

7. Hidden complaints are common for OVO and rose during DANA's spikes: the text model flags 16.9% [15.0, 19.0] of OVO's 4–5★ reviews as complaints vs 1.2% [1.1, 1.3] for ShopeePay, and DANA's rate rose from 4.7% [4.1, 5.3] in the week of Sep 14 to 16.9% [16.0, 17.9] in the week of Sep 21; these are lower bounds, since the model finds about half of 4–5★ complaints on gold (49% [31, 66]) (phase7_hidden_complaints.png).

8. The main specific complaint differs by app: money lost, missing or taken without consent is the largest specific topic everywhere and is most frequent for OVO (32.8% [31.4, 34.3] of its unique non-short complaint texts), though this broad topic fit only 3 of 8 rows in the fresh fit check (1 partly, 4 no), while DANA Cicil defines DANA (10.2% [9.9, 10.6]) and loans/paylater stand out for ShopeePay (9.9% [8.9, 10.9]) and GoPay (7.9% [7.4, 8.4]) (phase7_topic_trends.png).

Secondary (versions): DANA v2.145 shows 68.8% [68.0, 69.7] complaints, but 72% of its reviews were posted on the three spike days; excluding those days it is 30.3% [28.8, 31.9] vs 20.8% [19.9, 21.6] for v2.139, so the version gap is mostly spike timing, not a version effect (phase7_versions.png).

External corroboration

News reports matched to the spike days (each source checked by hand). A match only shows that the spike coincides with a reported event; it does not show that the event caused the reviews.

• GoPay, Jul 28: coincides with a reported GoPay outage of about 4 hours from about 12:03 WIB (CNBC Indonesia, Bisnis, Liputan6, Metro TV).
• DANA, Jul 20: coincides with the reported final reward-redemption day of DANA's "Festival Bola Dunia" event (DANA Blog) and with a reported DANA disruption that day (Tribunsumsel), the same day "Reward points cannot be redeemed" gained share (finding 3).
• DANA, Sep 25: coincides with reported user complaints about DANA disruptions to transactions and app access (Indozone).
• DANA, Oct 1: no report was found for this spike.

Figures

Limitations

• Topic fit is modest. On a fresh 60-row check, 40% [26, 55] of fitted documents fit their topic ("yes"); specific topics 52%, the two broad topics 29%. Topic statements are weaker than complaint-share statements.
• Short reviews carry no reliable topic. Topic trends use fitted documents only (unique, non-short texts); short or duplicate rows assigned by transform fit their topic in 15% [5, 36] of cases and are excluded.
• Star-rule false positives. "Star 1–2 OR model" counts 1–2★ praise as complaints (4 of the 20 transform rows in the fresh check were 1–2★ praise). Gold precision is 0.973, so the effect is small but not zero; gold recall is 0.876, so complaint shares are slightly understated overall and 4–5★ complaints are understated most.
• Denominators differ by measure. Complaint shares use all reviews, duplicates included (33.1% overall; 51.4% on unique review texts, because duplicates are mostly short praise). Topic shares use unique, non-short complaint texts. Hidden complaints use all 4–5★ reviews.
• Spike weeks dilute topic shares. In a spike week the denominator jumps, so steady topics drop in share without fewer complaints (DANA Cicil falls to 3.4% in the week of Sep 21 vs 23% the week before).
• Partial and lagged data. The Sep 28 week has 5.7 days (the feed lags about 24 h) and is compared per day; the Jun 29 week is dropped. Dates are UTC.
• Versions are installed versions, not release timing, and new versions can coincide with spike days (see the secondary finding).
• Reviewers self-select. People who write a Play Store review (often right after a problem) are not a random sample of users, so complaint shares describe reviews, not the share of users with problems.
• No causal claims. Spike causes are inferred from review text only; read reports/phase7_spike_examples.csv (10 random complaint texts per spike week for the top 2 rising topics) before naming a cause.

Known failure modes of the model

Observed when testing the demo (scores are softmax outputs, not calibrated probabilities):

The second case is a complaint written with 4–5 stars, the group where the model is weakest: on the gold set it caught only 49% [31, 66] of complaints written with 4–5★ (reweighted).

A friendly opener before "tapi" (but) hides the complaint. A small check (report only; the model was not changed) scored five short complaints written for the test, with and without an opener (CPU, same preprocessing):

So the flip comes mainly from the positive opener followed by "tapi", not from the rating alone; rating talk in the text pushes the scores further in its direction ("bintang 5" toward positive, "bintang 1" toward negative). A likely cause is the weak training labels: reviews that praise first and complain after "tapi" often carry 4–5 stars, so they were labeled positive. Five constructed examples are not a measurement of how often this happens in real reviews.

English is out of scope. About 0.5% of the training reviews are English; English text can be misread completely (see worst app ever). The demo shows a warning when it detects English (lingua restricted to Indonesian, Malay and English).