All projects

01.2 Pensions and long-term investing · Pre-registered

Keeping a pension promise funded

In plain English

A pension plan owes its retirees payments for decades. What does it cost to keep that promise, and can an AI manage the money better than the rule most plans follow?

  1. The promise A pension plan has promised to pay its retirees every month for the next 60 years, and every payment rises with prices. Today it holds 90% of what those promises are worth. Whoever stands behind the plan has to make up the rest.
  2. The risk Stocks can close that gap, or blow it open in a crash. Bonds built to match the promise keep the gap where it is. Prices rising faster than anyone expected widen it, whatever the plan holds.
  3. Two ways to decide Most plans follow a rule called a glide path: hold more stocks while the plan is short, fewer as it fills up. I trained a small neural network that starts as that rule and learns where to leave it, by living through thousands of simulated futures.
  4. The honest result Across 10,000 simulated futures the learned policy cost $5.0 for every $100 promised, against $9.9 for the glide path. Across 34 real 20-year periods it cost $6.3 against $6.3: the test written in advance calls that no detectable difference. So I took the policy apart to find out why, and most of its simulated edge came from patterns the simulator repeats, which no real plan can count on.
Promised to retirees$100 In the fund today90% the gap the sponsor has to close 50%200%400%todayyear 5year 10year 15year 20 100% fully funded each line is one simulated future 60% stocks, 40% bonds A glide path: fewer stocks as it fills The learned policy In simulation: $5.0 learned, $9.9 glide path, per $100 promisedSimulation: $5.0 learned, $9.9 glide path On real history: $6.3 learned, $6.3 glide path

What it costs to keep a pension promise funded for 20 years, and whether a policy learned by simulation does better than the rule plans use. In simulation it does; on real history it is not detectably different from that rule.

What it provesIn 10,000 simulated futures the learned policy costs the sponsor $5.0 per $100 promised against $9.9 for a glide path, and $30.2 against $69.1 in the worst 5%: better by the pre-registered test. When yields never return to an average it is better than the glide path, and over 34 real 20-year periods it is not detectably different from it.

A pension plan that is 90% funded has promised inflation-indexed payments for 60 years. Its liability is valued off the 10-year yield with expected inflation; its assets are stocks, 10-year bonds and bonds matched to the promise. Four standard policies (60/40, liability-driven, fully matched, a glide path) are compared with a policy trained in PyTorch through a differentiable simulator of the plan, on what the plan costs its sponsor over 20 years and how bad that gets.

Stack Python, PyTorch, NumPy, pandas, pytest

10,000 simulated 20-year futures that training never saw. Cost to the sponsor per $100 promised; lower is better.
PolicyAverage costWorst 5%Fully funded at year 20Average in stocks
60/40$14.3$86.090%60%
Liability-driven 40/60$10.1$58.591%40%
Fully matched$18.2$62.457%0%
Glide path$9.9$69.193%31%
Learned policy$5.0$30.298%18%

Average of 10,000 simulated futures; a glide path costs $9.9. Against it the pre-registered verdict is better in simulation, better when yields never return to an average, and no detectable difference on real history.

The question

A pension plan is a promise to pay people for decades. The hard part is not earning a return. It is keeping enough money in the plan to honour the promise whatever markets, interest rates and prices do.

What a plan owes is its ; what it holds against it, as a share, is its . When the plan is short, its has to pay in the difference. So the question a pension fund lives with is not "which portfolio grows fastest?" but "what will this promise cost to keep, and how bad can that get?" This project builds a small pension plan, measures that cost under the policies plans really use, and asks whether a policy learned by simulation does any better.

The plan

  • The promise. Monthly payments for 60 years, each a little smaller than the last (a mature plan running off), and each paid in the money of its day: fully indexed to consumer prices. Its is 14.2 years at a 4% yield.
  • What it is worth. Future payments are projected with expected inflation and discounted at the 10-year government yield. When yields fall, the promise costs more; when prices rise faster than expected, it costs more again.
  • Where it starts. 90% funded: ten cents short on every dollar promised.
  • What it can hold. Stocks, 10-year government bonds, and that move with the promise when yields move. No borrowing, no betting against anything, and every trade costs 10 .
  • Who pays. Whenever the plan is short, the sponsor pays in a tenth of the shortfall a year. A surplus is never paid back. The cost of a future is everything the sponsor pays in over 20 years plus whatever is still missing at the end, per $100 promised, in today's money.

Five policies

  • 60/40. 60% stocks, 40% bonds, rebalanced every month. The classic, and it ignores the promise entirely.
  • Liability-driven 40/60. 40% stocks, 60% matching bonds: most of the fund moves with the promise.
  • Fully matched. Everything in matching bonds. The gap stops moving with interest rates, and the sponsor pays it off.
  • . 70% stocks while the plan is badly short, sliding to 20% once it is comfortably funded. The rule many plans actually run.
  • Learned policy. A small neural network (1,507 parameters) that starts as the glide path and learns where to leave it. It is trained in PyTorch by : its weights set the holdings, the holdings set the plan's assets, the assets set what the sponsor pays, and the gradient of that cost flows all the way back. It sees only what a plan would know that month, never a forecast.

The test, fixed first

Every rule below was committed to the repository before a single test result existed.

  • One score. A policy's average cost plus its average cost in the worst 5% of futures (the ). The network is trained on it and judged on it.
  • Test 1: 10,000 simulated futures. Each rearranges what stocks, yields and prices actually did over 882 months of US history. Training never saw the test futures.
  • Test 1b: futures where yields never come back. In test 1, yields are always pulled back toward their long-run average, and a network can learn to bet on that. So the same policies are run again, without retraining, on futures where yields simply drift.
  • Test 2: 34 real 20-year periods. A plan starts every December from 1972 to 2005 and lives through what actually happened next. The network used for each start was trained only on data from before it ().
  • The decision rule. If the whole 95% interval for the difference in score is below zero, the learned policy is better; if it is above, worse; otherwise there is no detectable difference. The full text is the protocol.

Results

In simulation the learned policy costs the sponsor $5.0 per $100 promised against $9.9 for the glide path, and $30.2 against $69.1 when it goes badly. Verdict: better.

10,000 simulated 20-year futures that training never saw. Cost to the sponsor per $100 promised; lower is better.
PolicyAverage costWorst 5%Fully funded at year 20Average in stocks
60/40$14.3$86.090%60%
Liability-driven 40/60$10.1$58.591%40%
Fully matched$18.2$62.457%0%
Glide path$9.9$69.193%31%
Learned policy$5.0$30.298%18%
Score of the learned policy minus each rule's (average cost plus worst-5% cost, per $100 promised), simulated futures. Below zero is better.
ComparisonDifference95% intervalPre-registered verdict
Learned policy against glide path−43.8−46.9 to −40.7better
Learned policy against 60/40−65.1−68.5 to −61.8better
Learned policy against liability-driven 40/60−33.5−35.7 to −31.4better
Learned policy against fully matched−45.5−47.3 to −43.7better

When yields never come back

Without retraining, on futures in which the yield just drifts: $7.0 against $14.1 on average, $49.7 against $113.5 in the worst 5%. Verdict against the glide path: better.

The same policies, not retrained, on 10,000 futures in which the yield never returns to an average.
PolicyAverage costWorst 5%Fully funded at year 20Average in stocks
60/40$18.4$121.185%60%
Liability-driven 40/60$14.2$95.985%40%
Fully matched$20.6$78.053%0%
Glide path$14.1$113.588%32%
Learned policy$7.0$49.792%17%

On real history

Across 34 real 20-year periods the learned policy cost $6.3 per $100 promised against $6.3 for the glide path and $9.7 for 60/40. Verdict against the glide path: no detectable difference. These periods overlap, so they amount to about three independent ones: enough to see a policy fail, not enough to prove a small edge.

34 real 20-year periods, starting each December from 1972 to 2005. They overlap: about three are independent.
PolicyAverage costWorst periodFully funded at year 20Average in stocks
60/40$9.7$53.9100%60%
Liability-driven 40/60$6.8$42.1100%40%
Fully matched$8.4$50.688%0%
Glide path$6.3$46.9100%28%
Learned policy$6.3$55.6100%17%

What it learned

It holds less in stocks than the glide path whether the plan is short or funded. Above all it follows the yield: the higher the yield, the more it puts in bonds matched to the promise.

A lower cost says nothing about how it was earned, so after the test the scored policy was taken apart. While the plan is under 80% funded it keeps 41% in stocks, where the glide path keeps 70%; once the plan is over 110% funded it keeps 16%, where the glide path keeps 20%. The yield moves it more. For plans between 85% and 100% funded it keeps 54% in stocks when the 10-year yield is under 3% and 5% when it is over 10%; the glide path keeps 49% either way. Almost everything else goes into , and its target weights move 69% of the fund a year: this is a policy that trades.

0%25%50%75%100%60%80%100%120%140%fully fundedLearned, yield at 3%Glide pathLearned, yield at 6%Learned, yield at 9%funded ratio when the decision is made
Share of the fund in stocks as the funded ratio alone changes: every other input is held at a middling value, halfway through the 20 years. The glide path looks at nothing else; the learned policy also looks at the yield, so it is drawn at three levels of it.
What the learned policy held in the test futures, by how well funded the plan was at each monthly decision, next to what the glide path holds at the same funded ratio. A description made after the test; it changes no verdict.
Funded ratioShare of monthsGlide path: stocksLearned: stocksLearned: 10-year bondsLearned: matching bonds
Under 70%1%70%49%2%49%
70% to 80%2%70%37%2%61%
80% to 90%7%59%26%2%72%
90% to 100%12%45%21%1%77%
100% to 110%10%28%18%2%80%
110% to 120%9%20%16%2%82%
120% to 150%20%20%15%3%81%
Over 150%39%20%16%3%81%
The same, by the level of the 10-year yield, for plans between 85% and 100% funded.
10-year yieldShare of those monthsGlide path: stocksLearned: stocksLearned: 10-year bondsLearned: matching bonds
Under 3%8%49%54%1%45%
3% to 5%18%48%37%1%62%
5% to 7%28%48%24%2%75%
7% to 10%35%49%12%2%87%
Over 10%11%49%5%2%94%

Is it the network, or just a better-tuned rule?

The glide path in the test has textbook numbers. So its four numbers were trained on the same simulations, by the same method and for the same score as the network. Tuned, it keeps 32% in stocks until the plan is 63% funded and slides to 22% by 79%. It costs $10.9 on average and $53.5 in the worst 5%, against $5.0 and $30.2 for the network and $9.9 and $69.1 for the untuned rule. Tuning alone gets 33% of the network's improvement in the score.

The network still costs less than the tuned rule in simulation, so tuning is not the whole story.

Where the saving came from

The simulated futures are built by replaying history in runs that average twelve months, and the network sees the last twelve months of returns, yields and prices. So the same policies were run, without retraining, on futures whose months are drawn one at a time: nothing in the recent past says anything about what comes next.

The same policies, not retrained, on 10,000 futures whose months are drawn one at a time instead of in runs of history. Cost to the sponsor per $100 promised.
PolicyAverage costWorst 5%Fully funded at year 20Average in stocks
Glide path$10.4$79.095%30%
Tuned glide path$11.7$59.986%23%
Learned policy$10.2$59.090%19%

There the network costs $10.2 on average, the glide path $10.4 and the tuned rule $11.7. Its lead over the tuned rule falls from −29.2 points of score to −2.4: 8% of what it was.

Most of the simulated saving came from patterns in the last twelve months that hold inside the runs of history the simulator replays. A plan cannot rely on them: real history, where every network had seen only the past, gave the same answer.

Everything in this section was done after the test, on the same networks (retrained from the same seeds and checked to reproduce the scored result to the last digit). None of it was pre-registered and none of it changes a verdict.

When the Fed surprises

The same rise in rates can hurt or help a pension plan, depending on why it happened.

This is where the three studies meet. World Shock Lab measures, across 287 to 334 Fed announcements, what a surprise in US monetary policy does to markets. Applied to this plan: a 10 policy surprise moves stocks −0.98% and the 10-year yield +3.4 bp on the day, which changes the value of the promise by −0.48%. For a 60/40 plan the stocks cost 0.53 points of and the cheaper promise gives 0.34 back: a net of −0.20. An information surprise of the same size, the Fed revealing good news about the economy, sends stocks and yields up together, and the same plan moves +0.62. A plan held entirely in does not move in either case.

The move in the plan's funded ratio, in points, on the day of a 10 bp surprise in US monetary policy, at 90% funded. In brackets: a band that adds the two estimates' 90% half-widths.
Surprise (+10 bp)Stocks10-year yieldThe promise60/40Liability-driven 40/60Fully matchedGlide pathLearned policy
Fed policy surprise−0.98%+3.4 bp−0.48%−0.20 (±0.37)−0.18 (±0.22)0.00 (±0.00)−0.24 (±0.29)−0.13 (±0.15)
Fed information surprise+0.41%+4.1 bp−0.57%+0.62 (±0.79)+0.35 (±0.46)0.00 (±0.00)+0.47 (±0.61)+0.25 (±0.32)

Both moves are inside their bands: the two channels are of similar size, so the net is small and these estimates do not settle its sign. A first-order calculation at the plan's starting position, not a forecast and not part of the pre-registered tests. Open the lab to see the same surprise spread across 53 markets.

How to read it

  • Tests 1 and 1b are run inside a model. The network was trained on the same kind of simulated future it is tested on. A good score there shows it learned the model well. It does not show the model is the world, and the section above shows which part of the model it leaned on.
  • History is the check on that. The walk through real 20-year periods uses networks that never saw the years they were tested on. Its verdict is the one to weigh most, and it is reported whichever way it went.
  • The baselines are simple on purpose. A real fund's glide path is tuned by people who know their plan. Outscoring a textbook rule in simulation is evidence the method works, not that it would outdo a pension fund's investment team.
  • What transfers. The parts a pension fund cares about are not the network: valuing a promise off the yield curve, measuring what moves the funded ratio, costing a policy in the sponsor's terms, and fixing the test before running it.

Limits

Pre-registered in git before the test ran. A stylised plan: one yield for all maturities, expected inflation is the last ten years' average, no inflation-linked bonds, no credit or private assets, no longevity risk. Simulated futures rearrange US history from 1953 and are the model the learned policy was trained on: against the glide path the verdict there is better (score difference −43.8, 95% interval −46.9 to −40.7). On futures in which yields never return to an average, without retraining: $7.0 against $14.1, better. On 34 real 20-year periods, which overlap and amount to about three independent ones: $6.3 against $6.3, no detectable difference. US data only.

  • The matching bonds hedge the promise's value perfectly when yields move, which no real portfolio does: a real curve twists, and real liabilities are uncertain.
  • Expected inflation is just the last ten years' average. It moves slowly and can be wrong for a long time, as it was in the 1970s.
  • The discount rate is a government yield. Many plans use a higher one, which makes the same promise look smaller and the plan look better funded.
  • Members never join, leave or live longer than expected, and the sponsor always pays.

Reproduce

The study lives in the Deep Portfolio Constructor repository, beside the wealth study it shares a method with. One command downloads the public data, trains the 24 networks and writes every number on this page:

python -m dpc.cli pension

The pension code has 30 tests of its own. The mathematics is checked against closed forms: a fully matched, fully funded plan stays exactly funded whatever yields do; the liability's value matches discounting each payment one by one; no decision depends on anything that comes after it.