Modern-Era Trend Following in Global Futures
Separating opportunity decay, signal conversion, and implementation in a 75-market futures study, 1990–2026.
Research question
Does systematic trend following continue to generate economically meaningful returns in modern futures markets, and how has its performance changed across time, market structure, volatility regimes, and implementation assumptions?
Why this matters
Trend following is one of the most widely documented systematic return sources, and also one of the most widely declared dead. Both claims are usually made without separating the three things that can change: whether prices still trend, whether diversification still reduces portfolio risk, and whether the cost of trading has grown relative to the signal. Distinguishing them matters for anyone deciding how much capital a trend program should receive.
Hypotheses
- 01H1: The sign of an asset's trailing 12-month excess return predicts the sign of its next-month excess return across a diversified futures universe.
- 02H2: The magnitude relationship is weaker than the sign relationship, implying most of the value is directional rather than proportional.
- 03H3: Predictability measured after 2008 is lower than predictability measured before 2009, and the difference is larger than sampling error alone would suggest.
- 04H4: Multi-horizon signal blending improves risk-adjusted results relative to a single 12-month horizon.
- 05H5: Correlation-aware allocation improves risk-adjusted results relative to equal weighting.
- 06H6: Realistic transaction costs reduce but do not eliminate the measured result at the turnover levels implied by monthly rebalancing.
Methodology
Signal definition
The canonical baseline is the sign of the trailing 12-month excess return, evaluated at month end and held for one month. Extended specifications blend fast, medium, and slow lookbacks so that horizon is treated as an explicit research variable rather than a fixed setting.
Portfolio construction
Two allocation schemes are compared: equal risk weighting across markets, and a correlation-aware allocation that reduces the weight of markets that are close substitutes. Both are constrained to a common target portfolio volatility so that comparisons are not driven by leverage.
Position sizing
Positions are scaled inversely to an estimate of each market's recent volatility, so that a fixed risk budget is expressed in each market regardless of its nominal price level.
Rebalancing frequency
Monthly rebalancing at month end is the primary specification. Weekly and quarterly alternatives are tested as robustness, and reported with their turnover so that cost is comparable.
Volatility estimation
Volatility is estimated from trailing daily returns with a stated window, lagged so that no information from the position-sizing day itself is used. Alternative estimators appear in the robustness section.
Return construction
Returns are built from continuous series constructed from individual contracts with a documented roll rule, converted to a single base currency where the contract is denominated otherwise.
Transaction-cost assumptions
Costs are applied per unit of traded exposure using a per-market estimate covering commission and spread, and results are reported gross and net. A cost-sensitivity table shows the assumption at which the conclusion would change.
Statistical tests
Sign and magnitude regressions use non-overlapping monthly returns; inference uses Newey–West standard errors and two-way clustering by date and market. Multiple-market tests carry a Holm correction, and headline statistics carry bootstrap confidence intervals.
Benchmark selection
The 12-month time-series momentum baseline serves as the internal benchmark for extended specifications, alongside a long-only diversified futures reference. [EXTERNAL BENCHMARK TO CONFIRM]
In-sample and out-of-sample decisions
The pre-2009 sample is used to fix specification choices; the post-2008 sample is reserved for evaluation. Any later change to the specification is documented as such rather than silently applied to the full history.
Data
Markets and instruments
A frozen universe of 75 global futures spanning equity index, government bond, currency, energy, metal, and agricultural groups, selected from 112 candidate Norgate continuous series on history and metadata sufficiency before any strategy performance was examined.
Sample period
January 1990 through August 2026. Strategy returns begin after the required lookback and volatility warm-up; funded portfolio comparisons use completed months through July 2026.
Frequency
Daily settlement data aggregated to monthly evaluation points.
Data source
Norgate Data futures package: individual delivery histories plus back-adjusted and unadjusted continuous contracts. Licensed vendor data; raw licensed files are not redistributed on this site or in any public repository.
Cleaning and validation
- Calendar alignment across exchanges and holidays
- Outlier screening against contract-level price limits
- Cross-checks between contract volume, open interest, and roll dates
- Duplicate and stale-quote detection
Missing-data treatment
Missing settlements are carried forward only within a stated tolerance; longer gaps remove the market from that period's universe rather than being interpolated.
Survivorship considerations
Delisted and discontinued contracts are retained for the periods in which they traded, so that the universe reflects what was investable at the time.
Contract-roll methodology
Roll dates come from the active contract identifier in the Norgate continuous series rather than an assumed annual frequency. The audit covers 13,875 forward transitions, of which 958 skip at least one listed delivery; both execution legs are charged, while the adjacent-contract price spread remains futures carry embedded in the return series.
Currency conversion
Non-base-currency contracts are converted at the corresponding daily rate, with the conversion applied consistently to both the return and the risk estimate.
Licensing restrictions
Licensed data cannot be republished. The repository documents the exact schema and transformation steps so another researcher can reproduce the workflow with an equivalent authorized dataset.
Key results
No results published yet
Benchmark gross Sharpe
—
1.64 before 2009 → 0.52 after 2008
Benchmark net Sharpe
—
1.51 before 2009 → 0.42 after 2008
12-month continuation coefficient
—
0.128 before 2009 → 0.038 after 2008
Multi-horizon book net Sharpe (post-2008)
—
0.62, versus 0.42 for the benchmark
Maximum drawdown (post-2008)
—
12.8%, versus 18.8% for the benchmark
Annual cost drag (post-2008)
—
0.75%, versus 1.20% for the benchmark
Funded simulation (74 markets)
—
Sharpe above cash 0.64; maximum drawdown 9.1%
Full-sample multi-horizon book
—
7.51% net CAGR, 9.08% volatility, 0.84 net Sharpe
Planned exhibits
- Canonical 12-month time-series momentum benchmark
- Multi-horizon trend portfolio
- Equal-weight allocation
- Correlation-aware allocation
- Pre-2009 versus post-2008 evidence
- Turnover and transaction costs
- Market and asset-class attribution
- Robustness and limitations
Every published chart will carry a descriptive title, axis labels, units, legend where needed, its sample period, a gross or net label, a short written interpretation, and accessible colors with tooltips.
Interpretation
What the tests can show
Whether directional predictability is present in the sample, whether it is weaker in the later subperiod, and how much of the gross result survives a stated cost assumption.
What I would infer
I do not read the evidence as saying that trend following has stopped working. A weaker continuation premium leaves less room for avoidable implementation loss and less justification for concentrated model choice. The claim that modern markets simply flip more often and punish every flip more severely is not supported; the narrower reading is that weaker continuation and less efficient paths reduce the fraction of large moves captured by the monthly rule.
What remains uncertain
Whether any measured change reflects a durable structural shift or an unusually adverse stretch of a stable process. The sample cannot cleanly separate these.
Robustness checks
- Alternative lookback windows
- Alternative holding periods
- Weekly and quarterly rebalancing
- Alternative volatility estimators
- Rolling and fixed subperiod tests
- Pre-2009 versus post-2008 comparison
- Asset-class exclusion tests
- Cost sensitivity across a range of per-market assumptions
- Parameter sensitivity surfaces
- Bootstrap inference on headline statistics
- Holm multiple-testing correction
- Alternative continuous-contract construction
Limitations
- Licensed data cannot be redistributed, which limits direct third-party replication.
- Continuous-contract construction is a modelling choice that affects measured returns.
- Cost estimates are assumptions, not observed fills; realized slippage may differ by market and period.
- Multiple specifications are examined, so nominal significance overstates evidence without correction.
- Capacity and market impact are not modelled at institutional size.
- Financing and margin treatment are simplified.
- Subperiod tests have low statistical power by construction.
Conclusion
The 12-month continuation premium is smaller after 2008, so the benchmark begins with a lower gross return, and costs consume a larger share of that thinner opportunity. Two forms of diversification improve how the residual premium is harvested: combining distinct definitions of a one-year trend, and combining lookback and holding periods so the whole book does not reverse on one rebalance date. The larger and more reliable improvement comes from horizon diversification: once targets are combined before execution, turnover and drawdown fall together. A correlation-only optimizer does not improve on equal weighting. The evidence favors simple diversification across trend definitions and signal vintages, internal trade netting, and explicit liquidity controls rather than parameter selection or additional covariance optimization.
Reproducibility
Repository structure
[REPOSITORY LAYOUT TO CONFIRM] Separate directories for data ingestion, signal construction, portfolio construction, statistics, and figure generation.
Environment
A pinned dependency file records exact package versions used to produce every published output.
Running the analysis
A single entry point regenerates every table and figure from cached intermediate data, given an authorized dataset in the documented schema.
Random seeds
Bootstrap procedures use fixed seeds recorded alongside each published interval.
Known replication differences
Differences in vendor data, roll conventions, and volatility windows will produce small deviations. Expected tolerances are documented per exhibit.
Downloads and links
Citation
This is working research, not a peer-reviewed publication. If you refer to it, please cite it as work in progress and note the status shown above.
BibTeX
@misc{bang_trend_following_modern_era,
author = {Bang, Pratik},
title = {Modern-Era Trend Following in Global Futures},
year = {2026},
note = {Status: Completed. Working research, subject to revision.},
url = {[PROJECT URL TO ADD]}
}Plain text
Bang, P. (2026). Modern-Era Trend Following in Global Futures. Working research (Completed). [PROJECT URL TO ADD]