Google search volume → next-week index direction

Search Signal Desk

When people search more for a topic than they did over recent weeks, the model goes short the index for the following week; when they search less, it goes long. This is the strategy from Preis, Moat & Stanley (2013), extended to topics as in Curme et al. (2014).

Loading live data…
This week's call
—

Track record of this exact setup
Before split
—
After split
—
Buy & hold, after split
—
—

Cumulative return of the strategy

The shaded band is where 95% of random weekly long/short strategies would finish on the same weeks. A strategy that stays inside the band is indistinguishable from coin flips.

Search strategy Buy & hold Random strategies, 95% range Out-of-sample split

Weekly search signal

Search volume versus its trailing Δt-week average. Above zero means a short call for the following week; below zero means long.

Searches up → shortSearches down → long

Is the edge decaying?

Strategy return over the trailing 52 trades. The article notes the relationship weakened after 2012; this shows whether that holds for your setting.

Trailing 52-week strategy return

Robustness across look-back windows

Full-period z-score for each topic basket at each Δt. A real effect should hold across neighbouring windows, not appear at one Δt only. |z| ≥ 1.96 is significant at 5%.

Every signal, ranked before the split

Picking the best performer after the fact is the classic trap with this kind of data. Ranked by return before the split, then judged on what happened after it. Select a row to load it.

Live call log

Each Monday the app records the call for every topic basket at Δt = 3 on the Dow and never changes it afterwards. Unlike the backtests above, this is a record of calls made before the outcome was known.

How the model works

  1. For each search week t, take search volume n(t) and its average over the previous Δt weeks, N(t−1, Δt).
  2. For a topic basket, average each term's relative change, so a high-volume word cannot drown out the others.
  3. If searches rose, sell the index at the close on the first trading day after the search week and buy back a week later. If they fell, do the reverse.
  4. Score it against random long/short strategies on the same weeks: z = return ÷ √(Σ weekly log return²).
Δ(t) = [n(t) − N(t−1,Δt)] / N(t−1,Δt)
position(t) = −sign(Δ(t))
r(t) = position(t) · [ln p(t+1) − ln p(t)]

Weeks where volume exactly equals its trailing average are skipped. On the original data, the single word “debt” at Δt = 3 returns +298% with this rule; the paper reports +326%, which falls inside the +282% to +338% range produced by the different ways those tie weeks can be broken.

What to keep in mind

  • Returns ignore trading costs, slippage and the cost of shorting. A weekly flip strategy pays these 52 times a year.
  • Google Trends rescales each term to 0–100 per download, and the current week is partial until Sunday. The nowcast line uses that partial week and can change.
  • The topic baskets here are hand-grouped approximations of the study's 55 Wikipedia-derived topics.
  • This is a research tool, not a recommendation. A signal with no out-of-sample edge should not be traded.

Sources