TimesFM 2.5 in the browser

Forecasting this week's news

Two kinds of current-affairs forecast, both running in your browser on the same model.

⚾ Who wins the World Series?

The same recipe as the World Cup forecaster:

1. Team strength: an Elo rating for every MLB team, updated after every game of the season and playoffs. Winners gain points and losers lose them, with a bonus for home field and bigger wins.
2. TimesFM reads each remaining team's rating history and forecasts its strength for the next 10 games, with an uncertainty band.
3. Simulation: the rest of the playoffs are played out 100,000 times, starting from the current series scores, with 2-3-2 home field. Each game is a weighted coin flip from the two teams' strengths.

What it doesn't know

• Starting pitchers, injuries, bullpen fatigue and travel. In baseball, the day's starting pitcher matters a lot.
• TimesFM's band on team strength is narrow (about ±10 rating points), so most of the answer comes from the season's results and the series score. TimesFM adds a little extra uncertainty.
• Sanity check: in MLB history, teams leading a best-of-7 series 2–0 have gone on to win about 84% of the time, and only one team (the 2004 Red Sox) has come back from 0–3.

📰 Reading the news: a warning light, not a crystal ball

TimesFM only reads numbers, so a small sentiment model reads the headlines for it. Every day, up to 100 Google News headlines per topic (oil, Bitcoin, each team) are scored with VADER, a lightweight sentiment model, and averaged into a daily news tone from −10 (very negative) to +10 (very positive).

• Tone is not direction. "War escalates" is negative news but pushes oil up. So the tone never moves the forecast line.
• ⚠ Alert: when the latest day's tone is unusually far from the topic's own normal (more than 2× its usual day-to-day swing), the page raises a warning: the pattern forecast may be about to break.
• Stormy-news band: on an alert, the forecast chart adds a wider amber range showing how far that price moved over the same number of days in its most turbulent past periods.

Would it have warned us about the oil spike?

We replayed the daily oil-news tone from August to October 2026, using only each day's headlines (no hindsight), against the Brent price:

Three of five alerts came just before big moves, and it missed one (+8% on 15 Sep). That's two months of data, so it's encouraging but not proof. And even the stormy band was too narrow for this war spike: only about 27% of Brent's real prices stayed inside it. News can tell you when to worry, not always how much.

🛢️ Oil and ₿ Bitcoin: the next 30 days

Oil is high because of the Iran war, and markets are waiting for the US Federal Reserve decision on 28 October. Click a card to forecast the next 30 days in the Forecast tab.

Read these forecasts with care

• Prices behave like a random walk. "No change" is a hard baseline. Tested on the latest 30 days, TimesFM roughly ties it: a few percent better on some days, a few percent worse on others. The useful part is the shaded band: how far prices could plausibly move.
• The model can't read the news. When we hid Brent's last 30 days (to 6 Oct 2026), the real prices stayed inside the 80% band only about 30% of the time: the war-driven spike broke the pattern. For Bitcoin the band held 80–90% of the real prices (it shifts a little as each day's data arrives), close to the 80% it is designed for.
• Public oil prices lag about a week behind the markets, so the forecast starts from the last published day.
• Tick Hold back in the Forecast tab to repeat this check yourself.

Sources: Brent (FRED DCOILBRENTEU), WTI (FRED DCOILWTICO), Bitcoin (FRED CBBTCUSD, Coinbase). News context: Yahoo Finance, 8 Oct 2026; Democracy Now headlines, 13 Oct 2026; MLB.com playoff schedule.

Forecast your own data in 5 minutes

TimesFM takes one list of numbers in time order and predicts the next ones. It was never trained on your data. It's a "foundation model" that learned general patterns (trends, seasons, cycles) from billions of time points, so you can try it on almost anything that's measured regularly.

1. What makes a good series

• One number per time step, such as daily sales, weekly rainfall, monthly bills or yearly harvest.
• Evenly spaced. Every step should be the same length: every day, every month or every year. Don't mix daily and weekly values.
• Oldest first, newest last.
• Gaps are OK. Leave a blank or write NA. The model treats it as missing; it doesn't treat it as zero.
• Units don't matter. Rupees, tonnes, °C and % all work, because the model rescales everything internally.
• More history helps. Aim for at least 2–3 full seasons, for example 3 years of monthly data if there's a yearly pattern. 30 points is the practical minimum.

2. Get the numbers into the page

Pasted text can look like any of these:

120, 135, 128, 150, 162
120 135 128 150 162
120
135
NA
150

3. Pick context and horizon

• Context is how many recent points the model reads. More is usually better, up to about 1–2k. For the sunspots example, use 1024+ so it sees several 11-year cycles.
• Horizon is how many steps ahead to predict. Keep it shorter than your history; accuracy falls the further out you go.

Which model: 8-bit or 4-bit?

8-bit (261 MB) gives almost the same answers as Google's full-precision model. 4-bit (146 MB) downloads faster and is fine on long histories (hundreds of points), but on short series (under ~64 points) its forecast can drift noticeably. If your data is short, use 8-bit. Either way, the browser caches the file after the first load.

4. Check whether it's any good

Leave Hold back last N points ticked. The page hides your most recent N values from the model, forecasts them, then draws the real values in orange. Two numbers appear under the chart:

• MAE vs actual is the average miss, in your units.
• MAE of "repeat last value" is the miss from just copying the last known number forward. If TimesFM's MAE is lower, it's adding real value. If it's not, your series may be mostly noise or driven by outside events.

When you're happy with the result, untick it to forecast the real future.

5. Read the chart

• The blue line is the median forecast: the model's single best guess.
• The shaded band covers the 10th–90th percentile. The model expects about 8 in 10 future values to land inside it. A wide band means "I'm unsure". That's often the most useful part.

6. Know the limits

• It only sees the numbers. It doesn't know about holidays, policy changes, wars or new COVID variants unless they show up as patterns in the history.
• It forecasts one series at a time. It can't use rainfall to predict crop prices, for example.
• For decisions that matter, compare it with simple baselines and domain knowledge. This is a demo, not financial or medical advice.

Practice datasets

These are made-up numbers built to look realistic, so you can practise each kind of pattern. Click one to paste it into the forecast tab.

Where to find real data

• World Bank Open Data: poverty, health, agriculture and education by country, mostly yearly. Click Download → CSV.
• FRED: 800k+ economic series, including crop and food prices. Use Download → CSV on any chart.
• Our World in Data: COVID-19, energy, CO₂ and farming. Each chart has a Download tab.
• FAOSTAT: crop production, yields and food prices.
• data.gov.in / data.gov: government open data, including rainfall, mandi prices and hospital admissions.
• Your own life: step counts, electricity meter readings, shop sales, website visits or milk yield from your cows.

If a download has many countries or products in one file, filter to one of them before using a column.