Could we have seen it coming?
I've been following Malaysia's EV market, and wanted to know whether social media could predict which cars would sell. My first version used machine learning on thousands of comments. The second version went back to basics: check the model, fix the data, and test a forecast honestly. This is what held up and what didn't, and why it matters for anyone who forecasts promotions, launches and competitors.
Can social media predict which EVs will sell in Malaysia? What held up, what didn't, and why it matters if you forecast promotions and launches.
EVs went from 1 in 100 new cars to about 1 in 10
Battery EVs as a share of all new car registrations, by quarter. The local Proton e.MAS 5 took over in 2026, just as the tax-free window for imported EVs closed.
Battery EVs as a share of all new cars, by quarter.
A complex model did worse than a simple rule of thumb
My first version used machine learning to pick out 110 themes in the comments and predict sales from them. I compared it with the simplest forecast there is: next month will be the same as this month. The simple rule was about three times more accurate.
A machine-learning model built on comments, against one rule: next month = this month. The rule was three times more accurate.
In your business: check every forecast beats last month or last year. Most forecast reviews skip this.check every forecast beats last month or last year.
The second version fixed the messy data
Two Chery models are sold in petrol and electric versions under the same name. The first version counted both, so 43% of its “EV sales” were petrol cars. The second version keeps electric registrations only.
Two Chery models share a name across petrol and electric. The first version counted both: 43% of its “EV sales” were petrol.
In your business: shipments or sell-out? Gross or net of returns? The target column is a decision. Get it wrong and every model downstream is wrong.shipments or sell-out? Gross or net of returns? Choose the target on purpose.
Events explain the big moves. Comments show how buyers took them. Nothing predicts them.Events explain the big moves. Comments don't predict them.
I dated 28 launches, price cuts and competitor moves, then checked each against what normally happens on random dates. Pick a model, then an event.
28 dated events, each checked against normal swings. Pick a model, then an event.
A forecast frozen at October 2025, scored against the next 10 months
A plain statistical baseline, then human adjustments with a reason and a source for each one. “Locked” means neither could be changed once the real figures came in.
A statistical baseline plus human adjustments. “Locked” means neither could change once real figures arrived.
- STEP 1Only what was known thenRegistrations up to October 2025, and news published before 1 November 2025. Nothing later.Data to Oct 2025, news before 1 Nov 2025.
- STEP 2Rules set before the resultThe method was picked by a rule fixed in advance: whichever was most accurate on past data. Every adjustment was written down with its reason and source.Method chosen by a fixed rule. Every adjustment has a source.
- STEP 3Freeze, then scoreAll 10 months of forecast numbers were saved with a digital fingerprint. Only then were the real figures loaded. No re-running, no tweaking.Fingerprinted, then scored. No re-running.
I ran this in September 2026, as if it were October 2025. The fingerprints prove the numbers weren't changed after scoring, and the source dates show exactly what evidence each adjustment used.
Run in Sep 2026 as if it were Oct 2025.
The tax-deadline rush: caught
Budget 2026 ended tax-free imported EVs on 31 December. Two adjustments (+18% in November, +36% in December) cited the news. Error on those two months fell from 35% to 15%.
Two adjustments for the Budget 2026 deadline cut error from 35% to 15%.
The Proton e.MAS 5: missed
Adjustment M6 assumed the new RM60k Proton would add about 600 cars a month, the pace of Proton's last launch. It sold 3,068 in January and 4,770 in August. The adjustment log shows exactly which assumption broke, so the next cycle knows what to fix.
Assumed about 600 a month. It sold 3,068 in January. The log shows which assumption broke.
The whole scorecard, all 10 months
Average miss as a share of actual sales. Lower is better.
- Same as last month
- 56%
- Baseline
- 56%
- With adjustments
- 52%
- Same as last month
- 31%
- Baseline
- 22%
- With adjustments
- 23%
Over the full test, the adjustments helped the model-level forecast a little and the market total not at all. That's the honest result. Where the evidence was specific, they cut error sharply. When an assumption broke, the log showed exactly which one.
Adjustments helped models a little and the market total not at all. They worked where the evidence was specific.
Swap “EV” for your category
Every driver here has a twin in FMCG and retail.
| In the EV study | In your business |
|---|---|
| Price cuts | Promotions and price-offs |
| Model launches | New product launches |
| Competitor launches | Competitor activity |
| Tax-deadline rush | Loading before a price or excise increase |
| Comment reaction | Retailer and consumer feedback |
Curious how this would work on your numbers?
I use the same approach with FMCG and retail teams: a baseline you can check, adjustments with a reason and an owner, and one locked final that everyone works from.
The same approach, for FMCG and retail teams.
Data: DOSM vehicle registrations via data.gov.my; public Facebook and YouTube comments, used in aggregate only.