The Strategy: Equity Momentum Rotation for TSP

What the Strategy Does

The model picks the single strongest TSP equity fund — C (S&P 500), S (small/mid-cap), or I (international) — and puts 100% of the balance there. Every Wednesday, it re-evaluates which fund has the best momentum and rotates if the leader has changed.

That's it. No bonds. No blended allocations. No F Fund, no G Fund. The entire thesis is that over a 15-to-20-year federal career, concentrating in the highest-momentum equity fund and staying invested through volatility produces better outcomes than any diversified or defensive approach.

This is a max-return strategy. It accepts higher drawdowns in exchange for higher terminal wealth. If watching your balance drop 30% during a crash would cause you to panic and override the system, this approach is not for you — and that's a perfectly reasonable conclusion.

How the 3-Way Ensemble Works

Rather than relying on a single lookback window, the model uses three separate momentum calculations that each "vote" on which fund to hold:

• Short-term signal (21/63-day): Compares 21-day momentum against a 63-day baseline. Captures fast trend changes.
• Medium-term signal (42/63-day): A slightly slower cross, reducing whipsaw from day-to-day noise.
• Long-term signal (126-day): Pure 6-month lookback. Captures secular trends and filters out short-lived rallies.

The fund that wins the majority vote (2 out of 3, or unanimous) gets the allocation. When the three timeframes agree, the signal is strong. When they disagree, the model defaults to C Fund — the S&P 500 — because historically, staying in large-cap U.S. equities beats sitting in cash or bonds during uncertain periods.

This ensemble approach was chosen after testing dozens of single-window and dual-window configurations. No single lookback period works well across all market environments. The 3-way vote smooths transitions and reduces unnecessary trades.

Why C Fund Is the Default (Not G Fund)

Most TSP allocation advice follows a simple pattern: when things look bad, move to G Fund. The model tested this approach extensively and rejected it.

A G-default version of the strategy — same momentum signals, but parking in G Fund when signals disagree — produced a Sharpe ratio of 0.97 (slightly better risk-adjusted) but only reached $666,000 over 15.7 years. The C-default version reached $1,158,000. That's nearly $500,000 left on the table for a 13% reduction in maximum drawdown.

The math is clear: the cost of safety is enormous over long time horizons. Every day spent in G Fund earning 2-3% is a day not compounding at 10-16% in equities. For a federal employee with 20+ years until retirement, the strategy accepts higher short-term pain for significantly higher terminal wealth.

The model spends 0% of its time in G Fund. The -33.7% maximum drawdown is the price of the $1.16M outcome.

The Absolute Momentum Filter

There is one safety mechanism: the absolute momentum filter. If the top-ranked equity fund has negative momentum (its return over the lookback period is below zero), the model will not hold it. In that rare scenario, capital moves to C Fund as the default rather than chasing a fund that's actively declining.

Notably, the model does not use a "G hurdle" — it doesn't require equity momentum to exceed the G Fund yield. That test was run, and requiring equity funds to beat G Fund's return caused too many false exits during choppy markets. The bar is simply: is the fund going up at all?

The 4 AI Research Agents

The momentum signal drives all allocation decisions. But numbers don't tell the whole story. Four AI research agents run weekly analysis to provide context around the quantitative signal:

• Macro Analyst: Analyzes Fed policy, CPI/PCE inflation data, employment reports (NFP, JOLTS), yield curve shape, and geopolitical risks. Outputs a Risk-On / Neutral / Risk-Off rating with three supporting bullet points. Uses the Global Macro-Economic Analyst methodology from the quant skill cloud.
• Sentiment Analyst: Tracks VIX levels and term structure, put/call ratios, AAII investor sentiment surveys, and institutional fund flows. Produces a Sentiment Score from -100 to +100 and flags contrarian signals when extreme readings suggest the crowd is wrong.
• Risk Manager: Monitors portfolio VaR (95% and 99% confidence), inter-fund correlations, volatility regime shifts, and tail risk indicators. When all equity funds start moving in lockstep (correlation above 0.7), diversification fails — the Risk Manager flags these conditions and estimates worst-case scenarios.
• Strategy Arbiter: Observes the entire debate between the other three agents, then synthesizes their analyses — including disagreements — into the weekly dashboard commentary. The Arbiter doesn't smooth over conflicts; it presents both sides.

How the Agents Communicate: The Debate Process

The agents don't simply run in isolation. The weekly analysis follows a two-round process:

1. Round 1 — Initial Analysis: Macro, Sentiment, and Risk agents run sequentially. Each receives the prior agents' output, so the Sentiment Analyst knows the Macro rating before forming a view, and the Risk Manager sees both before assessing portfolio exposure.
2. Round 2 — Debate: All three agents see each other's complete analysis. Each can file dissents (specific disagreements with named colleagues), agreements (reinforcing points), and flag risks the group is underweighting. This is a professional debate — if the Macro Analyst says "Risk-On" but the Risk Manager sees elevated VaR, that tension is surfaced explicitly.
3. Synthesis: The Arbiter reviews the full debate record and writes the weekly commentary. Dissenting opinions appear on the dashboard marked as such, so you see where the agents disagree and can judge for yourself.

This matters because sanitized consensus hides useful information. If the Macro Analyst is bullish but the Risk Manager flags a correlation crisis, you should see both perspectives — not just a blended "moderately optimistic" average.

Agent Memory and Self-Learning

Each agent maintains persistent memory across weeks:

• Prediction tracking: Every weekly prediction is recorded with the market context at the time. The following week, predictions are graded against actual market returns.
• Rolling accuracy: Each agent has a 4-week, 13-week, and 52-week accuracy score. These are displayed on the dashboard so you can see which agents are hot and which are cold.
• Auto-generated lessons: When an agent misses 3 out of 4 recent predictions, the system generates a lesson describing the pattern (e.g., "missed during Bear-Volatile regimes when market went up"). These lessons are injected into future prompts so the agent can self-correct.

The agents do not override the momentum signal. They provide context, dissent, and risk assessment — but the allocation follows the quantitative model that was backtested. Think of them as a research team that adds color to a systematic strategy.

Backtest Results: 2010-2026

The strategy was backtested over 15.7 years using actual TSP fund price data. Starting with a $100,000 balance and no contributions:

Metric	Strategy	Buy & Hold (Equal Weight)
Total Return	1,059%	266%
Annualized Return	16.9%	~8.5%
Final Value ($100k start)	$1,158,751	$366,000
Sharpe Ratio	0.95	~0.55
Sortino Ratio	1.20	--
Max Drawdown	-33.7%	~-30%
Alpha vs. Buy & Hold	+793%	--

The strategy turned $100,000 into over $1.15 million while buy-and-hold reached roughly $366,000. The alpha — 793 percentage points of excess return — came from three sources: staying in equities (C-default), concentrating in the best-performing fund, and Wednesday rebalancing which captures cleaner mid-week momentum signals.

Important: Backtests are not predictions. They show how a strategy would have performed using historical data. Markets change. The future 15 years will not look like the past 15. The strategy could underperform in a sustained bear market or a decade of flat returns. Use this data to understand the approach, not as a guarantee.

Monte Carlo Validation: Is This Luck?

A backtest can look great by accident — maybe the strategy just happened to hold the right fund at the right time. To test this, we ran a Monte Carlo permutation test: 10,000 simulations where a random fund (C, S, or I) was selected at each Wednesday rebalance instead of the momentum winner.

If the momentum signal is just noise, random selection should produce similar results. It didn't — not even close.

Metric	Momentum Strategy	Random Selection (mean)	Random 99th Percentile
Sharpe Ratio	0.956	0.672	0.850
Total Return	1,069%	462%	~780%

The strategy beat 100% of 10,000 random simulations on both Sharpe ratio and total return. The p-value is less than 0.0001 — the momentum signal is adding real, statistically significant value. It's not luck.

To put this in context: the best random simulation out of 10,000 achieved a Sharpe of 0.850. The strategy's 0.956 sits well above even this extreme outlier. The probability of a random strategy matching the momentum signal's performance is effectively zero.

The strategy holds C Fund 49.5% of the time, S Fund 32.4%, and I Fund 18.1%. It's not a disguised buy-and-hold — it actively rotates between funds, and that rotation is where the alpha comes from.

What Was Tested and Rejected

Dozens of variations were tested before landing on the current approach. Some notable ones that didn't make the cut:

• G-Fund default: Better Sharpe (0.97 vs 0.91) but $357,000 less in final value. Not worth the tradeoff on a 20-year horizon.
• Volatility-adjusted scoring: Over-promoted bonds and F Fund, reducing equity exposure during the periods that mattered most.
• Danger exit (5% drop in 10 days): Cut max drawdown to -29% but cost $137,000 in missed recoveries.
• Trailing stop (-15%): Generated false triggers during normal corrections, destroying returns.
• S-Fund tilt: Small caps underperformed large caps in recent years, hurting overall results.
• Mutual Fund Window: Fees killed any alpha at $100k account sizes. Full analysis here.

Every drawdown-reduction idea that was tested cost more in lost returns than it saved in reduced risk. The conclusion: for a max-return strategy, the best defense is staying invested.

Why Wednesday? Rebalance Timing

The model checks every Wednesday. This isn't arbitrary — it was tested against every day of the week and every biweekly schedule:

Rebalance Day	Final Value ($100k)	Sharpe
Wednesday	$1,158,751	0.95
Thursday	$990,686	0.90
Friday	$964,570	0.89
Tuesday	$937,672	0.88
Monday	$926,184	0.87

Mid-week rebalancing avoids Monday gap risk (weekend news) and Friday noise (options expiration, end-of-week positioning). By Wednesday, markets have digested the week's economic data and momentum signals are cleaner. This is a well-documented effect in momentum literature.

Biweekly schedules were also tested: 2nd & 4th Wednesday ($1.1M) beat 1st & 3rd ($951k) by $145k. But checking every Wednesday and only acting when the signal changes outperformed all biweekly schedules — most weeks require no transfer at all.

TSP Transfer Rules and the 2/Month Limit

TSP allows two interfund transfers (IFTs) per calendar month. After the second, you can only move money into the G Fund. The strategy is designed around this constraint:

• Weekly Wednesday check: The model evaluates every Wednesday but only executes a transfer when the momentum leader changes. Most weeks, no action is needed.
• Transfer budget enforced: Only 7% of months hit the 2-IFT limit. The strategy averages 0.84 transfers per month — well within budget.
• No G Fund moves: Since the strategy never allocates to G, it never triggers the "G-only after 2 IFTs" rule.

When the limit is hit, the model simply holds the current fund until the next month resets the IFT budget. In backtesting, this cost less than 1% of total return over 15 years.

Key distinction: Contribution allocation changes are free and unlimited. They do not count as IFTs. Changing where your future contributions go costs nothing. Only moving your existing balance between funds counts toward the 2/month limit. More on contributions here.

A Note on the AI Agents

The AI "agent debate" feature is interesting, but be clear about what it is: analytical commentary layered on top of the momentum signal. The agents do not override the momentum signal. They provide context, dissent, and risk assessment. Don't confuse the sophisticated-sounding analysis with the actual allocation driver, which is purely mechanical.

The agents can unanimously disagree with the momentum signal — as they did during the March 2026 selloff — and the allocation still follows the model. This is by design. The backtest validates the momentum engine, not the agent commentary. The agents exist to help you understand why the model is positioned where it is and what risks it might not be capturing.

Who This Strategy Is For

This is a thoughtfully built tool for federal employees who want a systematic, high-risk/high-reward TSP strategy. The underlying logic is sound. But it's optimized for a specific type of investor:

• Long time horizon — 15+ years until retirement. The longer the horizon, the more drawdowns are absorbed by compounding.
• High risk tolerance — you can watch your balance drop 30%+ and not panic-sell or override the system.
• Discipline — the hardest part isn't understanding the strategy, it's following it during drawdowns. If you can't commit to that, the Balanced mode (Sharpe 1.18, max DD -17.6%) may be more appropriate.
• Systematic mindset — you prefer rules-based decisions over gut feelings and market predictions.

If your priority is minimizing volatility or you're within 5 years of retirement, a more conservative approach — including the Balanced override or a target-date L Fund — may be more appropriate. There's no single right answer, only tradeoffs.

Read before you follow: This is a personal project tracking a quantitative strategy. It is not financial advice. Please read this full page and understand the drawdown risk, the mechanical nature of the allocation, and the limitations of backtesting before following any of its signals.