It means replacing opinion with a measurable rule. Rather than deciding “the market looks expensive to me”, you define in advance what you observe, what triggers a decision, and what you do when it happens. The rule is then applied to history to see how it would have behaved.

The difficulty is not finding a rule that worked in the past — that is trivial, and that is the trap. Search long enough and you will always find a rule that would have worked. It will not work tomorrow. Three requirements separate serious research from tinkering:

• The rule is written before it is tested. Thresholds, floors and ceilings are fixed first. They are not adjusted afterwards to flatter the result.
• It is tested on data nobody has looked at. You train on one period, validate on the next, lock it, and do not go back.
• The number of trials is counted. Test a hundred ideas and the best will look excellent by pure chance. Statistical tests correct for that effect — ours are published below, including when they disappoint.

And a fourth requirement, less technical: failures are kept. A research file containing only successes is not a research file.

Riviera TIDE was born from a field observation rather than an ambition for performance: our clients cope badly with deep drawdowns, and a deep drawdown costs years. The starting question was therefore not “how do we earn more” but “how do we lose less, without giving up most of the upside”.

The first version relied on a classic trend signal — a long moving average with a dead band meant to limit round trips. It was abandoned: too slow on the exit, and above all vulnerable to whipsaw. In a market without trend, the price oscillates either side of the average; each crossing triggers a trade devoid of informational content, of which only the cost remains.

The second version changed logic. Rather than following a price, it seeks to identify a change of market state — what academic research calls a regime change. The retained method is published in a peer-reviewed journal from a leading asset-management publisher; we re-implemented it independently, then cross-checked it against the authors’ reference code — a double control that reveals either our error or a divergence from the publication.

An internal review uncovered an execution bias in a complementary layer: a signal was being applied to the same day’s return instead of the next day’s. Once corrected, the layer lost all its value. It was removed, and an internal rule followed: independent recomputation is mandatory before any figure is published.

A published idea is not a strategy. Here is the path a piece of research must travel before we agree to put it into production — and the fact that it can be dropped at any stage.

Most of the approaches we evaluated did not survive. That is the normal outcome of a serious protocol, and it is the part of the file we find most instructive. Click an approach to see the figure that eliminated it.

The engine uses no macroeconomic data, no valuation, no news and no options. It reads only the daily returns of the index itself, and derives two measures from them: a short-term trend and a measure of dispersion. These feed an algorithm that divides history into a small number of persistent states.

The heart of the mechanism lies elsewhere: changing state has a cost. The algorithm agrees to switch only if the evidence is strong enough to pay that cost. That is what stops it reacting to every bad week. This level of inertia is not set by hand: it is automatically re-selected every two years, on past data alone.

On the way down, the trigger is a regime in which returns deteriorate and volatility rises durably — not a price level, not a moving-average crossover. It fires within a few days to a few weeks after a genuine break, and it deliberately does not fire on an isolated shock.

On the way up, it is the same test in reverse, with one added confirmation: exposure only rises again if the favourable state holds for fifteen consecutive sessions. The engine is therefore deliberately asymmetric: quick to leave, patient to return.

Let us be precise about what that costs. In simulation, the recovery after a low took between 60 and 209 sessions. A fast re-entry layer had been built to correct this: it did not survive honest execution accounting and was removed. It is the most significant limitation of the mechanism.

Finally, there is no top or bubble detector. The engine makes no claim to spot euphoria: its only bullish behaviour is that confirmed recovery.

• Universe. Six equity markets tracked in production, each assessed independently. In research, the same rule was transferred as-is to 34 underlyings.
• Frequency. Daily assessment after the close, switching only when the regime changes. Roughly 2.3 switches a year on the S&P 500.
• Execution. Signal computed on day D, executed on day D+1. Never the same day.
• Fallback. When a market is unfavourable, its sleeve moves to a defensive reserve — intermediate government bonds and gold — each filtered by its own long-term trend. If both are poorly oriented, the sleeve goes to cash.
• No short selling, no bet on a fall. Exposure never drops below the profile’s floor, and never turns negative.
• Human decision. The engine produces a proposed reallocation. The adviser validates it, the client consents to it. No order is executed automatically.
• Circuit breakers. No proposal if the signal is more than three business days old, or if an index switches more than eight times over a rolling twelve months — that would point to an anomaly, not to a market.
• Costs accounted for. Spread, slippage, switching fees, fund management fees and execution delays within assurance-vie are modelled in the simulations.
• Traceability. Raw data and signals are archived with a digital fingerprint, so they can be inspected later.

The validation protocol strictly separates training from testing: a rolling learning window, a setting re-selected periodically on past data only, a buffer period between learning and testing, delayed execution, and a battery of tests designed to detect any leakage of information from the future into the past.

The same rule, frozen, was then transferred to 34 underlyings that had never been used to build it. Honest verdict: it adds value on 14 of them. Not on all 34.

Finally, we corrected for the effect of the number of trials. The results are published as they stand:

• Probability of overfitting: 39%. In other words, choosing the “optimal” configuration is meaningless — and we do not do it.
• Reality check against a balanced benchmark: p = 0.088. Outperformance is not statistically demonstrated.

That is why we make no claim to beat the market. What the simulations support is a reduction in risk: over 2002–2026, a return comparable to that of the US market for a maximum drawdown roughly half as deep. That is a more modest result than a promise of return, and a far more solid one.

The other limitations, unvarnished:

• Returns remain below those of permanent exposure on the cautious profiles. That is the price of protection, not a flaw.
• The recovery after a low is slow: 60 to 209 sessions in simulation.
• In 2018, the engine did worse than the index. In 2022, it helped the most aggressive profiles only marginally.
• It does not work on sovereign bonds, nor for choosing between markets, nor on several emerging markets or commodities.
• The exposure floor, which protects against false signals, has a cost in a deep crash on the higher profiles.
• Live monitoring has only just begun. Everything above is simulation.

Riviera TIDE is a research and decision-support tool. It is not a promise of future performance, and it replaces neither your risk profile nor your adviser’s judgement.