Strategies · Climate

El Niño farm tilt

A strong El Niño shifts rain away from India, Southeast Asia and Australia, and harvests there tend to suffer. When NOAA’s index says the Pacific is strongly warm, put half the portfolio in farm commodities while it lasts.

Pre-registered Sep 26, 2026, before it was tested. Results computed Sep 26, 2026.

Open in the builder

Hypothetical backtest; not investment advice; paper trading only.Past results, real or simulated, do not predict future ones. ARCANE places no live orders from this page.

The rule

Trades on the first trading day of each month

While the latest Oceanic Niño Index is above +1.0 °C, hold half agricultural commodities (DBA) and half the S&P 500 (SPY); otherwise the S&P 500. Checked monthly.

Where the numbers came from. +1.0 °C is the usual line for a moderate-to-strong El Niño. Half and half was set by hand.

Hold what is below
If the Oceanic Niño Index is greater than 1
then: Split equally across 2 blocks
Hold DBA
Hold SPY
Otherwise
Hold SPY

As text

symphony "El Niño farm tilt"
description "Half in farm commodities while the Pacific is strongly warm."
rebalance monthly
benchmark SPY
weight equal
  if ARCANE:enso/oni > 1
    weight equal
      asset DBA
      asset SPY
  else
    asset SPY

How it would have done

Feb 2, 2007 to Sep 25, 2026, against SPY

The whole history

32 trades

A year
+9.7%
Annualized return
In all
+514%
Feb 2, 2007 to Sep 25, 2026
Sharpe
0.58
Return per unit of risk
Worst fall
−55%
Max drawdown
Swings
19%
Volatility, a year
Beta
0.97
To the benchmark

Each decision reads closing prices up to that day and trades at the next day’s close. It trades on the first trading day of each month. Costs: 5 bp of slippage on every trade and 0.3 bp of fees on sales. Uninvested cash earns nothing. Hypothetical: no real orders were placed.

$0$2$4$6$82008201120142017202020232026Held out

Growth of $1
SPY, held
Held-out period

−50%−25%0%

What it held: SPY 89% of the time; DBA + SPY 11% of the time.

In sample and the sealed holdout

The holdout was set before the preset was tested, and opened once

Measure	In sample Feb 2, 2007 – Sep 24, 2021	Sealed holdout Sep 27, 2021 – Sep 25, 2026
Annualized return	+9.1%	+11.4%
Sharpe ratio	0.54	0.71
Sortino ratio	0.75	1.03
Max drawdown	55.2%	24.5%
Calmar ratio	0.17	0.46
Volatility	20.0%	17.2%
Alpha against the benchmark	−0.7%	−1.3%
Beta	0.97	0.97
Correlation	0.99	0.97
Winning days	55%	54%
Turnover a year	0.2×	0.3×
Chance the Sharpe is above zero	98%	94%

Is it real?

Checked on the in-sample window only

Overfitting: Holds up
No clear edge over its benchmark

Its numbers survive being nudged and the count of every try.

After counting 3 backtests of this idea, the out-of-sample Sharpe still clears the bar luck would set (deflated Sharpe 1.00).
The best version in one half of history mostly holds up in the other (probability of overfitting 6%).
Against SPY the difference (−1.0% a year) is within what luck produces (9% chance it truly beats it).

Backtests counted	3
Deflated Sharpe ratio (0.95 is the house bar)	1.00
Probability of backtest overfitting (20% is the house ceiling)	6%
Walk-forward Sharpe (choosing, then testing)	0.85
Walk-forward Sharpe with costs doubled	0.85
Extra return a year against SPY	−1.0%

The nudged versions (2)

Change	Sharpe
threshold 1 on enso/oni → 0.5	0.47
threshold 1 on enso/oni → 1.5	0.58

The house backtest gate

It does not pass. A strategy needs every one of these before ARCANE would paper-trade it.

Check	Value	Needs	Result
Evaluation ran without problems	clean	no problems	Passed
Report belongs to this pre-registration	el-nino-farm-tilt	el-nino-farm-tilt	Passed
Evaluated only inside the pre-registered data window (holdout untouched)	2007-02-02..2021-09-24	within 2007-02-01..2021-09-24	Passed
Walk-forward folds	11.000	>= 3	Passed
Out-of-sample observations	2772.000	>= 252	Passed
Deflated Sharpe ratio (3 trials counted)	0.996	>= 0.95	Passed
Probability of backtest overfitting	0.064	<= 0.2	Passed
Out-of-sample Sharpe (annual, after costs)	0.854	>= 0.5	Passed
Out-of-sample maximum drawdown	0.337	<= 0.25	Failed
Sharpe with doubled costs	0.854	> 0 (needs a turnover series)	Passed
Sealed holdout: positive and not clearly worse than the backtest	mean 0.000486, consistent	mean > 0 and consistent	Passed

The pre-registration

Written before the backtest; changing any of it makes a new variant

Claim: Holding half DBA and half SPY while the Oceanic Niño Index is above +1.0 °C, and SPY otherwise, beats SPY buy-and-hold on Sharpe after 5 bp costs over the sealed holdout.
Why the edge should exist: Strong El Niño years cut yields in major growing regions; agricultural futures reprice as crop reports confirm the damage, months after the ocean data showed the risk.
Where it comes from: slow information
Who is on the other side: Commodity index holders and hedgers who wait for crop reports rather than ocean data.
Why it is not arbitraged away: Weather-driven commodity funds exist but are small; the signal fires in a few years a decade, too rarely for most mandates.
Universe: Oceanic Niño Index, DBA, SPY
In-sample window: Feb 1, 2007 to Sep 24, 2021
Sealed holdout: Sep 27, 2021 to Sep 25, 2026
Evaluation: walk-forward, on sharpe, against SPY buy and hold; 13 variants planned
Costs: 5 bp slippage a trade, no commission
Capacity: $250,000: DBA trades tens of millions of dollars a day; a quarter of a million is a small share.
What stops it: Paper Sharpe significantly below the backtest after 12 months of signal-on time. DBA closes or changes its index.

The data

Prices: Yahoo Finance (unofficial chart API): adjusted daily closes for DBA, SPY. Unofficial and not licensed for redistribution. A licensed vendor (Alpaca, Tiingo or Polygon) is a founder decision before the builder is sold.
Oceanic Niño Index: NOAA Climate Prediction Center, from Mar 15, 1950. History: revised. NOAA’s ONI table from 1950. A season counts as public on the 15th of the month after it ends (CPC publishes early in the month). NOAA revises the underlying sea temperatures, so history is today’s record; the official El Niño and La Niña episodes are decided after the fact and are not used.